Prescription drug expenditures in the United States: the effects of obesity, demographics, and new pharmaceutical products.

Datta, Anusua

1. Introduction

During the period 1990-1998, real per capita expenditures on prescription drugs in the U.S. increased by 84% (1996 dollars, GDP deflator). Beginning in 1993, the average annual percentage increase in prescription drug spending exceeded the overall percentage increase in national health expenditures. By 1998, the annual percentage growth in prescription drug spending was more than 16%, while overall spending on health care rose less than 6% (compared to the previous year). In total, U.S. consumers spent more than $90 billion on prescription drugs in 1998, or $334 on a per capita basis (Centers for Medicare and Medicaid Services 2002). Not surprisingly, prescription drug coverage and its associated costs have become issues in national as well as state political campaigns. (1)

But the rising expenditures were not simply caused by higher prescription drug prices. While the CPI (Consumer Price Index) for prescription drugs rose by 42%, nominal per capita expenditures on prescription drugs rose 120% during the period (U.S. Bureau of the Census 2002). Consequently, higher rates of prescription drug consumption explain at least part of the story. Further, while overall prescription drug spending has increased rapidly, per capita prescription drug use in the United States varies widely by state. In 1998, some states had per capita prescription drug spending of more than $400, while other states' residents spent half that much.

While we do not wish to absolve the pharmaceutical manufacturers, this data suggests that arguments placing the blame for high prescription drug costs on the prices and profits of pharmaceutical manufacturers may be missing a portion of the story. The literature on prescription drugs suggests a number of causes (e.g., new drug introductions, aging of the population, insurance coverage). However, Berndt (2002) notes that a shortcoming of the existing literature is the lack of quantitative estimates of the causes for increased consumption of prescription drugs since the mid-1990s. In a similar vein, Kane (1997) concludes that the effect of the managed care revolution has been difficult to separate out from the other forces affecting pharmaceuticals.

To better inform public policy, we provide some quantitative estimates of the factors that have contributed to the increase in prescription drug expenditures. We investigate the role of public health factors (obesity, smoking, and alcohol consumption rates); aging (population 65 and over); access to medical care (managed care enrollments and the relative size of the uninsured population); new pharmaceutical products; income; and unemployment on real prescription drug expenditures, using panel data from all 50 U.S. states for the period 1990-1998. The estimates should allow better projections of the anticipated increase in prescription drug expenditures.

Gauging the costs of smoking, obesity, and alcohol consumption is also important. High costs associated with any of these public health problems make it easier to justify costly government programs to reduce their prevalence. Thus, public policy responses to rising prescription drug expenditures will vary based on the source of the increase. To the extent that rising prescription drug expenditures are the result of rising real income rather than public health conditions, the case for laissez-faire is stronger.

We find that obesity rates, the relative size of the over-65 population, new pharmaceutical products, unemployment, and real income exert a significant effect on prescription drug expenditures. Overall, the estimates suggest that about 8% of the increase in spending on prescription drugs during the period 1990-1998 can be explained by the increase in obesity. In contrast, rising real incomes account for about 55% of the increase in prescription drug expenditures during the period. While the percentage of the population over 65 and new drug approvals exert a significant positive effect on per capita prescription drug expenditures, an increase in the unemployment rate reduces per capita prescription drug expenditures.

The paper is organized as follows: section 2 discusses the background literature and provides the motivation for this study; section 3 discusses the fixed effect instrumental variables model and data sources; in section 4, we present the estimates of the instrumental variables; and section 5 discusses the results from the prescription drug expenditure model. Section 6 concludes.

2. Background

A number of recent papers have explored the causes of rising prescription drug prices and expenditures (Berndt 2001; Reinhardt 2001; Thomas, Ritter, and Wallack 2001; Berndt 2002; Kaufman et al. 2002; Sturm 2002). (2) In general, the explanations have focused on the demand side and have identified five basic causes: prescription drug spending is a small percentage of total health expenditures; increases in the percentage of the population with insurance coverage for prescription drugs; the introduction of new blockbuster drugs; aging of the population; and public health factors.

Berndt (2001) and Reinhardt (2001) argue that cost-cutting efforts did not focus on prescription drugs because prescription drugs account for a small portion of total health expenditure (8% in 1998) and that increased third-party drug (insurance) coverage creates problems of moral hazard. Patients are more likely to use lower-priced generic products if they have to pay a large portion of costs out-of-pocket rather than when they are covered by a third party. Indeed, the percentage of out-of-pocket drug spending fell from 92% in 1965 to 26% in 1998, implying an increase in prescription drug use. Lundin (2000) shows that patients with large out-of-pocket costs are more likely to choose to use generic drugs. Apparently, consumers with full coverage have little incentive to search for low-cost alternatives to brand-name drugs. Purchases of low-cost generic drugs may also fall because of the introduction and aggressive marketing of new blockbuster drugs by the pharmaceutical industry (Berndt 2001, 2002). However, there is little systematic evidence to support this claim.

Of course, prescription drug spending may rise because of changes in the demographic composition of the population rather than consumption choices. Thomas, Ritter, and Wallack (2001) and Kaufman et al. (2002) provide data that shows that prescription drug usage is the highest among the elderly due to a higher incidence of cardiovascular and gastrointestinal diseases and chronic conditions. For example, among elderly people who spent more than $4000 annually on medications, 88% took cardiovascular medications, 64% took gastrointestinal medications, and 57% took lipid-lowering medications.

The rapid rise in obesity over the past decade has been a significant source of worry for public health officials (Nestle and Jacobson 2000). Obesity is associated with a variety of risk factors for cardiovascular diseases, such as hypertension, elevated cholesterol, and type-II diabetes, as well as cancer, stroke, and osteoarthritis and other diseases (Must et al. 1999). Using survey data from 1997-1998, Sturm (2002) measures medication costs by mapping survey responses on regularly used medications to insurance claims for prescription drugs and wholesale prices for other types of medication. He found that obesity increases an individual's average medication costs by 77% and smoking (present and past) increases such costs by 28-30%. However, the analysis fails to control for insurance status, income, or new drug introductions.

Other studies, such as Coulson and Stuart (1995) and Coulson et al. (1995), use subject self-reports of health status to control for health as a determinant of prescription drug use. However, such treatment of public health factors suggests that these factors are beyond the reach of public policy. Most of these studies are based on survey data or self-reports, which make it harder to generalize the results. Moreover, the studies do not provide statistical estimates of the extent to which each of these factors affects prescription drug expenditures.

Suraratdecha (1996) analyzes prescription drug spending across states for years 1980-1990 using four independent variables: the percentage of the population that is over 65 years of age, the proportion of Medicaid recipients, real income per capita, and physician services expenditures per capita (a proxy for the number of physicians per capita). Each of the variables, save physician services, has a significant positive impact on per capita prescription drug expenditures. However, Suraratdecha does not account for public health factors and new drug introductions.

In sum, the literature on prescription drug spending does not provide a comprehensive analysis of the factors underlying rising prescription drug costs. While some studies (Suraratdecha 1996; Lundin 2000; Berndt 2001; Reinhardt 2001) have estimated the effects of economic factors such as income and insurance coverage on prescription drug spending, they do not account for the role of public health factors. In addition, the analyses fail to correct for the two-way causality that exists between insurance factors and prescription drug expenditures. Other studies (Thomas, Ritter, and Wallack 2001; Kaufman et al. 2002; Sturm 2002), while highlighting the significance of public health factors in prescription drug spending, do not control for effects of economic variables such as insurance coverage and income. Finally, some of the latter estimates are based on survey response data, which makes it harder to generalize the results.

The purpose of the present study is to fill this gap in the literature by examining the role of public health factors, along with income and insurance status, on prescription drug expenditures. In addition, the study also accounts for the effect of new drug introductions, which is shown to be a significant factor. A fixed effects panel data model is used to control for state-specific differences. Our model also controls for the problem of endogeneity between prescription drug expenditures and some key explanatory variables through an instrumental variables approach.

3. Empirical Model and Data

Empirical studies of the determinants of prescription drug expenditures in the U.S. have been conducted using a cross-section framework (e.g., Lundin 2000; Sturm 2002), time-series (Berndt 2001, 2002), or a simple pooling of cross-section and time-series data (e.g., Suraratdecha 1996). However, assuming a common intercept with cross-section and time-series data ignores "individual effects," which can lead to biased results (Islam 1995). To eliminate such biases, we employ a fixed-effects panel-data model to analyze prescription drug expenditures across the United States for the period 1990-1998. F-test and Hausman test results suggest that in this case, the fixed-effects model is superior to the common intercept and random-effects model. (3)

The fixed-effects model assumes that differences across states are captured by differences in the constant term.

[Drug.sub.it] = [X'.sub.it][delta] + [[alpha].sub.i] + [u.sub.it], (1)

where i indexes states; t indexes year; [Drug.sub.it] represents real per capita prescription drug expenditure; [X'.sub.it] is a vector of explanatory variables; [[alpha].sub.i] is the time-invariant, unobserved state effects; and [u.sub.it] is the random error term that varies across states and time periods. Under the assumption that the [[alpha].sub.i] are constant and [u.sub.it] is normally distributed with a zero mean, Equation 1 is estimated using the least squares dummy variable (LSDV) method. The LSDV eliminates a major portion of the variation between the dependent and independent variables when the between-cross-section and between-time variation is large.

Based on the prior discussion, we regressed real per capita prescription drug expenditures (Drug) on per capita income (Income), unemployment rate (Unemployment), percentage of the population over 18 years of age that is obese (Obese), percentage of the total population over 65 years of age (65+ Years), percentage of the population without insurance (Uninsured), percentage of the population over 18 years of age that smokes (Smoke), per capita alcohol consumption (Alcohol), and the percentage of the total population that is enrolled in HMOs (HMO).

Measuring the effects of the introduction of new drugs by pharmaceutical companies is less straightforward. Following Cockburn (2004), we have included FDA approvals of new molecular entities (NMEs) as our measure of new drugs. We expect that an increase in any of the independent variables (except for Uninsured and Unemployment) would cause an increase in per capita prescription drug use. The Uninsured variable counts individuals with either private or public (e.g., Medicaid) insurance as insured. Unfortunately, we are not able to identify the percentage of the population that has insurance coverage for prescription drugs.

We include the unemployment rate (in addition to the percentage uninsured) because Ruhm (2003) finds that unemployment improves health. A one-percentage-point decrease in the unemployment rate significantly increases the prevalence of medical problems, acute morbidities, restricted-activity days, bed days, ischemic heart disease, and invertebral disc disorders. If higher unemployment rates cause health to improve, prescription drug expenditures may fall as well.

A closer look at the explanatory variables suggests that there may be an endogeneity problem between some key determinants of prescription drug expenditures and prescription drug expenditures. HMO, Uninsured, and 65+ Years may have two-way causality with prescription drug expenditures. HMOs have traditionally offered better coverage for prescription drugs than standard fee-for-service plans. Thus, higher rates of HMO enrollment may cause higher spending on prescription drugs. On the other hand, it is also possible that individuals react to higher prescription drug costs by joining HMOs.

Similarly, we would expect two-way causality between the prescription drug expenditures and the percentage of uninsured. Uninsured individuals may reduce purchases of prescription drugs because all costs are out of pocket, but higher prescription drug costs may make employers less able to offer insurance to employees. Likewise, an increase in the relative size of the 65+ population may increase per capita expenditures on prescription drugs, but an increase in prescription drug expenditures may raise life expectancy and increase the relative size of the 65+ population. Consequently, we addressed these endogeneity biases using an instrumental variables approach. We use the labor force participation rate as an instrument for HMO enrollments, the poverty rate as an instrument for the percentage uninsured, and birthrate as an instrument for percentage of the population 65 and over. We estimate the following set of equations:

[HMO.sub.it] = [Y'.sub.1it] [[gamma].sub.1] + [Z'.sub.it] [[delta].sub.1] + [[phi].sub.1i] + [e.sub.1it], (2)

[Uninsured.sub.it] = [Y'.sub.2it] [[gamma].sub.2] + [Z'.sub.it] [[delta].sub.2] + [[phi].sub.2i] + [e.sub.2it], (3)

65 + [Years.sub.it] = [Y'.sub.3it] [[gamma].sub.3] + [Z'.sub.it] [[delta].sub.3] + [[phi].sub.3i] + [e.sub.3it], (4)

where [Z'.sub.it] is a vector of exogenous variables (obesity rate, smoking rate, per capita income, unemployment rate, alcohol consumption, and NME), [Y'.sub.it] are the instruments--poverty rate, labor force participation, and birthrate--used in Equations 24, respectively, and [[phi].sub.i] and [e.sub.it] are the state-effects and regression errors for the respective instrumental variables equations. We then save the estimated values of the dependent variable for Equations 2, 3, and 4 for use in the equation for prescription drug expenditures:

[Drug.sub.it] = [[??].sub.1][HMO.sub.it] + [[??].sub.2] [Uninsured.sub.it] + [[??].sub.3]65 + [Years.sub.it] + [Z'.sub.it] [[delta].sub.4] + [[alpha].sub.it] + [u.sub.it]. (5)

Equations 2 through 5 were estimated using the instrumental variables procedure with fixed effects available in LIMDEP. The LIMDEP panel procedure for instrumental variables automatically corrects the standard errors to reflect the use of predicted variables as covariates.

Data for the study was collected from the Centers for Medicare and Medicaid Services website (prescription drug expenditures) and The Statistical Abstract of the U.S. (per capita income, percentage of the population over 65, poverty rate, percentage of population with insurance coverage, and birthrate). Real per capita prescription drug spending and real per capita income are in 1996 dollars and include both public and private spending. In the results reported in this paper, nominal values are converted using the GDP deflator. Converting the nominal values using the CPI had no effect on the results. Data on the employment variables are taken from the Bureau of Labor Statistics' Local Area Unemployment Statistics (unemployment rate and labor force participation).

The Centers for Disease Control and Prevention's Behavioral Risk Factor Surveillance System supplied the data on smoking and obesity rates. Obesity is based on Body Mass Index (BMI) where BMI is weight in kilograms divided by height in meters squared. An individual with a BMI [greater than or equal to] 30 is considered obese. The alcohol consumption data is from the National Institute on Alcohol Abuse at the National Institutes of Health. The data reports annual per capita consumption of alcohol by state based on sales data. All alcohol consumption (e.g., beer, wine, and spirits) is converted to an ethanol equivalent and the per capita calculation includes the population 14 years of age and older. InterStudy's Competitive Edge HMO Industry Report provides HMO enrollment rates. Data on New Molecular Entities was collected from the U.S. Food and Drug Administration Center for Drug Evaluation and Research. (4)

The data contains annual observations on each variable across the 50 U.S. states for each year during the period 1990-1998 (9 years and 50 cross-sections). While prescription drug expenditures grew steadily from 1990-1998, substantial differences among the states remained. In 1998, nominal per capita spending on prescription drugs was highest in New Jersey ($437), West Virginia ($428), Pennsylvania ($420), and Florida ($416) and lowest in Alaska ($217), New Mexico ($231), California, ($231), and Colorado ($244). Table 1 reports means, standard deviations, and definitions for the dependent and independent variables. There are some missing observations on smoking and obesity rates for Alaska, Arkansas, Kansas, Nevada, New Jersey, Rhode Island, and Wyoming (13 total missing observations). (5) Because the autoregressive panel procedure requires a balanced panel, we interpolate these values. (6) The District of Columbia is excluded because estimates of its HMO enrollment are not reliable.

4. Instrumental Variables Estimates

In this section, we discuss the results from the fixed-effects regressions on HMO, Uninsured, and 65+ Years. Because newly developed drugs may take a year or so to penetrate the market, we report parameter estimates for new drug approvals (NMEs) with no lag and a lag of two years. (7) The results reported in Table 2 provide some interesting insights. The estimated values of the dependent variables in Table 2 are used as instruments in fixed-effects regressions on real per capita prescription drug spending.

HMO Equations'

The results show that the rate of HMO enrollment is greater in high-income states and states with higher rates of obesity. A $1000 increase in per capita income causes an increase of about 2.5 percentage points in HMO enrollment. A one percentage-point increase in obesity causes an increase of about 0.75 percentage points in HMO enrollment. Each of these factors is consistent with the argument in Dranove (2000) that HMOs were a response to rising health care costs. In addition, HMO enrollments are lower where labor force participation is higher. A one percentage-point increase in labor force participation causes a decrease of about 1.1 percentage points in HMO enrollment. Higher unemployment rates also lower HMO enrollments. A one percentage-point increase in the unemployment rate lowers HMO enrollments by about half a percentage point.

Finally, the estimates show that new drug introductions initially have a significant positive effect on HMO enrollments. An additional NME initially raises HMO enrollments by about 0.08 percentage points. This is a substantial effect. In an average year, there are about 30 NMEs. Thus in an average year, HMO enrollments rise 2.4 percentage points because of new drug introductions.

Uninsured Equations

The percentage of uninsured people is positively related to both state per capita income and the poverty rate. A $1000 increase in per capita income causes an increase of about 0.5 percentage points in the percentage of uninsured. This likely reflects the higher health care costs that prevail in high-income areas. A one percentage-point increase in poverty causes an increase of 0.26 percentage points in the percentage of uninsured. Surprisingly, higher alcohol consumption lowers the percentage of uninsured. This may be the result of the link between alcohol consumption and depression. An increase of one gallon in the annual per capita consumption of alcohol lowers the number of uninsured by about two percentage points. Smoking, obesity, unemployment, and new drug approvals have no statistically significant impact on the percentage of uninsured. Even with a two-year lag, additional NMEs have no effect on the percentage that is uninsured.

65+ Years Equations

Increases in obesity, alcohol consumption, income, and the birthrate are associated with a reduction in the percentage of the population over 65. A one percentage-point increase in obesity reduces the relative size of the over-65 population by 0.01 percentage points. New drug introductions (lagged two years) cause a small but significant impact on the percentage of the population over 65. Thirty new drug introductions (the sample average) increase the percentage of the population over 65 by 0.12 percentage points. Per capita alcohol consumption, income, and the birthrate may be endogenous. While high alcohol consumption may lower life expectancies, it is also possible that individuals on average consume less alcohol as they age. While the over-65 population may migrate away from the high income/high tax states, it is also possible that a high percentage of the population over 65 decreases income, because income typically drops at retirement. Higher birthrates may reduce the relative size of the over-65 population, but a large over-65 population will reduce the birthrate.

5. Determinants of Rising Prescription Drug Expenditures

Table 3 shows fixed effects regression on real per capita prescription drug spending. Columns 1 and 2 of Table 3 show the results of a fixed-effect regression with per capita income, unemployment, percentage of the population over 65, percentage of population uninsured, smoking rate, obesity rate, per capita alcohol consumption, HMO enrollment rates, and new drug approvals (NMEs) as independent variables.

To test whether the instruments (poverty rate, labor force participation rate, and birthrate) are uncorrelated with the error term in the final prescription drug expenditures equation, we estimate the final prescription drug expenditures equation (with the fitted values for HMO, Uninsured, and 65+ Years) and obtain the residuals. Then, we regress the residuals on all exogenous variables (including the three instruments and a constant). The results show that for each of the instruments (birthrate, labor force participation, and poverty rate) the t-statistics in the residuals regression are insignificant, with t-values between 0.017 and 0.216 ([r.sup.2] = 0.002). Thus, the instruments are uncorrelated with the errors in the main equation (prescription drug expenditures).

Consequently, we employ poverty rate, labor force participation rate, and birthrate as instruments in the instrumental variables estimates for HMO enrollment and percentage of population uninsured, and percentage of the population over the age of 65. Columns 3, 4, and 5 of Table 3 substitute the estimated values of HMO enrollment and percentage of population uninsured, and percentage of the population over 65 for the actual values using the estimates in Table 2. As in Table 2, we report estimates for new drug approvals and new drug approvals lagged two years. Because there is some evidence of serial correlation, we ran an autoregressive procedure. The results of the autoregressive procedure with instrumental variables appear in column 5 of Table 3. In the discussion that follows, we use the estimates in column 5 as the definitive results unless indicated otherwise.

Among the public health indicators, obesity is positively related to prescription drug expenditures and is significant in all specifications. Obesity is associated with a variety of risk factors for cardiovascular disease, such as hypertension, elevated cholesterol, and type-II diabetes, as well as an increased risk of cancer, stroke, osteoarthritis, and other diseases (Must et al. 1999). These secondary effects of obesity typically require additional expensive medicines to treat complications and can substantially increase expenditures on prescription drugs. Surprisingly, neither smoking nor alcohol consumption has a significant effect on prescription drug expenditures. Of course, smoking rates may impact prescription drug expenditures after many years. We experimented with lags of zero to three years on the smoking variable and found that none of the specifications were significant. Ideally, we would have also investigated lags of 10 to 20 years, but we could not locate data to conduct such an analysis.

The estimates in Table 3 (column 5) show that a one percentage-point increase in the obesity rate raises per capita prescription drug expenditures by $1.63. While obesity exerts only a modest effect on per capita prescription drug expenditures, obesity increased dramatically from 1990-1998. The U.S obesity rate was 11.6% in 1990, but by 1998 it had risen to 18.3%, which is an increase of 57%. Thus, the increases in obesity rate from 1990-1998 raised per capita prescription drug expenditures about $11. (8) Overall, the estimates suggest that about 8% of the increase in spending on prescription drugs during the period 1990-1998 can be explained by the increase in obesity.

Our estimates suggest that obesity is associated with a 71% increase in prescription drug spending. To help picture this, suppose that one person in a population of 100 becomes obese. Because per capita spending rises by $1.63, the change in obesity status must have raised spending for that individual by $163. We know that the mean per capita prescription drug spending (1996 dollars) for our data set is $229 and 163/229 is 0.71. (9) Thus, according to our estimates, the effect of obesity on prescription drug expenditures is about the same as the estimates in Sturm (2002). Using a different data set (Medical Expenditure Panel Survey [MEPS]), Sturm found that obesity was associated with a 77% rise in prescription drug costs.

Turning to demographic factors, the results show that the percentage of the population over 65 is an important influence on prescription drug use. A one percentage-point increase in this population leads to an increase in real per capita prescription drug expenditures per capita of $21.86. The relative size of the over-65 population varies a great deal among states and it accounts for much of the variation across states at any point in time. While the magnitude of this effect is large, changes in percentage of the population over 65 cannot account for the rise in prescription drug expenditures for the period 1990-1998. During the period 1990-1998, the percentage of the population over 65 increased only 0.1 percentage points, from 12.6% to 12.7%. However, if the projected increase in the relative size of the population over 65 does in fact materialize, it would exert a strong effect on prescription drug spending.

Because prescription drugs are likely a normal good, we expect increases in income to raise prescription drug expenditures. Income sensitivity of prescription drugs likely increased as a result of introduction of a series of so-called "lifestyle" drugs (e.g., Rogaine, Viagra). Access to medical care also likely rises with income. The estimates in Table 3 show a strong positive effect of about $23 on real per capita prescription drug spending for every $1000 increase in real per capita income. During the period 1990-1998, real per capita income rose about $3550. Consequently, increases in income caused an increase in real per capita prescription drug spending of about $80. Overall, the estimates suggest that about 55% of the increase in spending on prescription drugs during the period 1990-1998 can be explained by the increase in real per capita income.

The effect of unemployment on per capita prescription drug expenditures was more modest. The estimates in Table 3 suggest that a one percentage-point increase in the unemployment rate decreases per capita prescription drug expenditures by about $4. During the period 1990-1998, the U.S. average annual unemployment rate rose from 5.6% in 1990 to 7.5% in 1992, and then fell to 4.5% in 1998. Thus, the net change in the unemployment rate accounts for only a small percentage of the increase (about 2%) in per capita prescription drug expenditures for the period 1990-1998.

Access to prescription drugs is measured by the percentage of the population enrolled in HMOs and the percentage of the population that is uninsured. Increases in the percentage of the population without health insurance were expected to decrease the rate of prescription drug use. People without health insurance must pay full price for their prescriptions, and as a result will tend to purchase less. Rather surprisingly, the uninsured coefficient is insignificant in both regressions. This may be the result of price discrimination or a failure of the variable to capture the extent of prescription drug coverage. Frank (2001) shows that individual cash payers at pharmacies pay roughly 30% more than those who use managed care plans for prescription drugs. Thus, higher prices per prescription may offset the reduction in the number of prescriptions. In addition, the percentage of uninsured may show no effect on prescription drug expenditures because medical insurance coverage for individuals may or may not include coverage for prescription drugs.

From a theoretical perspective, increases in HMO membership might either increase or decrease expenditures. Because HMOs tend to have good prescription drug plans, increases in the percentage of the population covered by HMOs should increase the rate of prescription drug use (higher Q). However, HMOs are also able to obtain prescription drugs at lower prices (Frank 2001). The estimates in column 5 of Table 3 show that an increase in HMO enrollment has no effect on prescription drug expenditures. While the estimates in columns 1 through 4 show a modest and consistent effect (a one percentage-point increase in the percentage of the population enrolled in HMOs raises per capita prescription drug expenditures by about $1.30), the correction for serial correlation eliminates this effect.

Finally, lagged new drug approvals (measured as New Molecular Entities approved by the FDA) show a stronger effect on per capita prescription drug expenditures than contemporary new drug approvals. An additional new drug approval (NME) lagged two years raises per capita prescription drug spending by about $0.65. We experimented with lags of zero to three years. For lags of zero and three years, we found no significant effect of new drug approvals on prescription drug expenditures. For lags of one and two years, the effect was significant and positive. We report only the two-year lag in Table 3 because the parameter estimate and the t-statistic were a bit higher for the two-year lag.

This result should be interpreted cautiously because measuring innovation as the number of NMEs does not account for differences in the medical or economic significance of each of the new molecules. Nevertheless, our estimate suggests that the effect of new drug approvals is substantial. On average, the FDA approved about 30 new drugs annually during the period 1988 to 1996. This implies that in an average year, per capita prescription drug spending rose by about $19 solely as a result of new drug introductions. The cumulative effect of the new drug introductions on per capita prescription drug expenditures is less certain as drugs approved in the current year may displace spending on older drugs.

6. Conclusion

During the period 1990-1998, real per capita expenditures on prescription drugs in the U.S. increased by 84% (1996 dollars, GDP deflator). Nominal per capita expenditures on prescription drugs rose 120% during the period while the CPI for prescription drugs rose by 42% (U.S. Bureau of the Census 2002). This suggests that higher rates of prescription drug use must be at least part of the story. This paper provides some quantitative measures of the factors that have contributed to the increase. We examined the factors driving prescription drug expenditures using panel data from all 50 U.S. states for the period 1990-1998. Results indicate that public health, population over the age of 65, new pharmaceutical products, and income are all important in explaining prescription drug expenditures.

Among public health indicators, obesity is a significant factor, but smoking and alcohol consumption are not. Overall, the estimates suggest that about 8% of the increase in spending on prescription drugs during the period 1990-1998 can be explained by the increase in obesity. Obesity rates registered a dramatic increase (57%) during this period. This suggests reductions in the obesity rate could yield modest reductions in health care spending, holding health outcomes constant. Indeed, we also find that increases in the obesity rate also raise HMO enrollments. A one percentage-point increase in obesity causes an increase of 0.8 percentage points in HMO enrollment.

Another factor that is strongly significant is the percentage of population over the age of 65. Although the percentages do not change much over the time period, it appears that the magnitude effect of the population over 65 is very large, making it a significant determinant of prescription drug expenditures. As other studies have shown, high incidence of chronic illnesses and poor health among the elderly population contribute to higher spending on prescription drugs.

The results also show that a large chunk of the increase in prescription drug expenditures is caused by rising income. Prescription drugs are normal goods and 1990-1998 marks a period of relative prosperity in the United States. The subsequent rise in real incomes accounts for about 55% of the increase in real per capita prescription drug spending. Despite the strong effect of income on prescription drug expenditures, we find no evidence that changes in the percentage of uninsured residents cause changes in prescription drug expenditures. However, the uninsured variable counts only the percentage of the population that has insurance coverage for medical care and insurance for medical care may fail to offer coverage for prescription drug costs. More surprising, labor force participation and unemployment have no statistically significant impact on the percentage of uninsured.

Finally, we find that new drug approvals (NMEs) cause substantial increases in per capita prescription drug expenditures. An additional new drug approval (lagged two years) raises per capita prescription drug expenditures by about $0.65. Given that in an average year about 30 new drugs are approved, the average annual increase in per capita prescription drug spending caused by new drug introductions is about $19. Moreover, we find that new drug approvals raise HMO enrollments and increase the relative size of the over-65 population (with a two-year lag).

We wish to thank Subarna Samanta, Judith Shinogle, and Karen Conway for helpful discussions and comments. Julie L. Hotchkiss provided us with valuable suggestions for an earlier version of this article, and we benefited from advice given by two anonymous referees at this journal. Ellie Fogarty, Ravi Kaneriya, Michael Ferlise, and Leigh Ann Culbertson provided valuable research assistance.

Received August 2004; accepted January 2006.

References

Berndt, Ernst. 2001. The U.S. pharmaceutical industry: Why major growth in times of cost containment? Health Affairs 20:100-14.

Berndt, Ernst. 2002. Pharmaceuticals in U.S. health care: Determinants of quantity and price. Journal of Economic Perspectives 16:45-66.

Bureau of Labor Statistics. 2003. "Local Area Unemployment Statistics." Accessed 8 February 2004. Available at http:// www.bls.gov/lau/home.htm.

Centers for Medicare and Medicaid Services. 2002. "1998 State Estimates--All Payers--Prescription Drugs." Accessed 10 September 2002. Available at http://cms.hhs.gov/statistics/nhe.

Cockburn, Iain M. 2004. The changing structure of the pharmaceutical industry. Health Affairs 23:10-22.

Coulson, N. Edward, and Bruce C. Stuart. 1995. Insurance choice and the demand for prescription drugs. Southern Economic Journal 61 : 1146-57.

Coulson, N. Edward, Joseph V. Terza, Cheryl A. Neslusan, and Bruce C. Stuart. 1995. Estimating the moral hazard effect of supplemental medical insurance in the demand for prescription drugs by the elderly. American Economic Review 85:122-6.

Danzon, Patricia M. 1997. Price discrimination for pharmaceuticals: Welfare effects in the U.S. and E.U. International Journal of the Economics of Business 4:301-21.

Dranove, David. 2000. The economic evolution of American health care. Princeton, NJ: Princeton University Press.

Elzinga, Kenneth G., and David E. Mills. 1997. The distribution and pricing of prescription drugs. International Journal of the Economics of Business 4:287-99.

Frank, Richard G. 2001. Prescription Drug Prices: Why do some pay more than others do? Health Affairs 20:115-28.

Hitt, Greg. 2002. Drug makers pour ad money into final days of campaign. The Wall Street Journal, 4 November, P. A1.

InterStudy Competitive Edge (Various years). HMO industry report. St. Paul, MN: InterStudy Publications.

Islam, Nazrul. 1995. Growth empirics: A panel data approach. Quarterly Journal of Economics 110:1127-70.

Kane, Nancy M. 1997. Pharmaceutical cost containment and innovation in the United States. Health Policy 41:S71-S89.

Kaufman, David W., Judith P. Kelly, Lynn Rosenberg, Theresa E. Anderson, and Allen A. Mitchell. 2002. Recent patterns of medication use in the ambulatory adult population of the United States. Journal of the American Medical Association 282:337-44.

LIMDEP version 8.0. 2002. Plainview, NY: Econometric Software, Inc.

Lundin, Douglas. 2000. Moral hazards in physician prescription behavior. Journal of Health Economics 19:639-62.

Must, Aviva, Jennifer Spadano, Eugenie H. Coakley, Alison E. Field, Graham Colditz, and William H. Dietz. 1999. The disease burden associated with overweight and obesity. Journal of the American Medical Association 282:1523-29.

National Center for Chronic Disease Prevention and Health Promotion, Centers for Disease Control and Prevention. 2002. "Behavioral Risk Factor Surveillance System." Accessed 10 September 2002. Available at http:// www.cdc.gov/brfss/index.htm.

National Institute on Alcohol Abuse and Alcoholism. 2004. "Apparent Per Capita Alcohol Consumption: National, State, and Regional Trends, 1970-2002." Accessed 8 February 2004. Available at http://www.niaaa.nih.gov/ databases/consum03.htm.

Nestle, Marion, and Michael Jacobson. 2000. Halting the obesity epidemic: A public health policy approach. Public Health Reports 115:12-24.

Reinhardt, Uwe E. 2001. Perspectives on the pharmaceutical industry. Health Affairs 20:136-49.

Ruhm, Chrostopher J. 2003. Good times make you sick. Journal of Health Economics 22:637-58.

Sturm, Roland. 2002. The effects of obesity smoking, and drinking on medical problems and costs. Health Affairs 21:245-53.

Suraratdecha, Chutima. 1996. A model of state-level prescription drug expenditures in the USA. Applied Economic Letters 3:289-92.

Thomas, Cindy P., Grant Ritter, and Stanley S. Wallack. 2001. Growth in prescription drug spending among insured elders. Health Affairs 20:265-77.

U.S. Bureau of the Census. (Various years). Statistical Abstract of the United States. Accessed 10 September 2002. Available at http://www.census.gov/statab/www/.

Van Vliet, Rene C. 2001. Effects of price and deductibles on medical care demand, estimated from survey data. Applied Economics 33:1515-24.

(1) Anticipating some form of new regulation, the pharmaceutical industry spent at least $16 million on issue ads in the 2002 election cycle (Hitt 2002). In particular, the industry was concerned about Democratic Party proposals that would administer prescription drug benefits through Medicare and about legislation that might make generic drugs more accessible. In the 2004 presidential campaign, prescription drug costs remained a significant issue as George W. Bush touted the creation of a prescription drug benefit under Medicare and as John Kerry argued for reimportation of prescription drugs from Canada and overhauling the Medicare drug plan.

(2) Two other strains of the literature consider the market factors that allow price discrimination in prescription drugs (Danzon 1997; Elzinga and Mills 1997; Frank 2001) and the effect of insurance on prescription drug use (Coulson and Stuart 1995; Coulson et al. 1995; Lundin 2000; Van Vliet 2001).

(3) The random-effects regressions yield similar results, but a Hausman test rejects the null of random effects.

(4) See also Cockburn (2004) for this data.

(5) The missing observations are: 1990 (Alaska, Arkansas, Kansas, Nevada, New Jersey, Wyoming), 1991 (Kansas, Nevada, Wyoming), 1992 (Arkansas, Wyoming), 1993 (Wyoming), and 1994 (Rhode Island).

(6) We interpolate these values by locating the most recent year for which we have an observation for state i. For that year, we calculate the difference between the state rate and the national rate (for smoking or obesity). We then use that difference to adjust the national rate and use it for the state value. For example, the obesity rate for Alaska in 1990 is missing. Alaska's obesity rate for 1991 is 13.4 and the U.S. obesity rate for 1991 is 12.6--a difference of 0.8. Since the U.S. obesity rate for 1990 is 11.6, the imputed value for Alaska in 1990 is 11.6 + 0.8 = 12.4. Repeating the analyses to exclude Wyoming (four missing values) or 1990 (six missing values) does not affect the results. To further reduce confusion, we use n = 450 throughout. The estimates in Table 2 and columns 14 in Table 3 are essentially the same regardless of whether we interpolate for the missing 13 observations on the obesity and smoke variables. Rather than switch between the interpolated data set (n = 450) and the noninterpolated data set (n = 437), we use the interpolated data throughout.

(7) We investigated lags of zero, one, two, and three years on the NME variable in the prescription drug equation. Lags of one and two years showed a significant effect on prescription drug expenditures. Lags of three and zero showed no statistically significant effect. We report estimates of NME for a lag of zero and two years. We report the zero lag to show that there is no contemporaneous effect and we report the two-year lag because the parameter estimate is a bit larger than the one-year lag.

(8) Calculation: (18.3 - 11.6 = 6.7 and 6.7 x $1.63 = $10.92).

(9) As a robustness check, we also ran both pooled regressions and between regressions estimates. In each case, the parameter estimate for obesity is positive, significant, and slightly higher than in the fixed-effects regressions reported in Table 3. These estimates show a weaker positive effect (but still significant) for income and percentage of the population over 65 years of age than the fixed-effects estimates. The HMO variable is not significant in either of these alternative specifications.

Donald Vandegrift, Professor of Economics, The College of New Jersey, 2000 Pennington Rd., Ewing, NJ 08628, USA; E-mail vandedon@tcnj.edu.

Anusua Datta, Associate Professor of Economics, School of Business Administration, Philadelphia University, School House Lane and Henry Avenue, Philadelphia, PA 19144, USA; E-mail dattaa@philau.edu, corresponding author.

Table 1. Descriptive Statistics

 Standard
Variable Mean Deviation Minimum Maximum

Drug 229.23 58.72 114.96 423.29
Obese 14.71 3.003 6.9 23.9
Smoke 23.2 2.88 13.2 31.7
Alcohol 2.31 0.541 1.2 4.78
Income 22.61 3.41 15.21 35.95
65+Years 12.65 2.05 3.97 18.55
HMO 16.22 11.96 0 54.2
Uninsured 14.01 4.09 2.1 25.6
Poverty 13.17 3.94 5.23 27.63
Labor 50.93 2.78 42.45 56.76
Unemp. 5.57 1.51 2.2 11.4
Birthrate 14.93 1.81 11 21.52
NME 30.67 9.77 22 53
NME (-2) 27.33 9.72 18 53

[Drug.sub.it]: Real per capita prescription drug expenditures
(1996 dollars) for state i in year t. [Obese.sub.it]: Percentage
of the population 18 years of age and over in state i who is
obese (BMI [greater than or equal to] 30) in year t.
[Smoke.sub.it],: Percentage of the population 18 years of
age and over in state i that smokes in year t. [Alcohol.sub.it]:
Per capita (population 14 years of age and older) annual
consumption of alcohol (ethanol equivalent) in gallons for
state i in year t. [Income.sub.it] Real per capita income
(in thousands of 1996 dollars) for state i in year t. 65+
[Years.sub.it]: Percentage of the total population 65 years
of age and older for state i in year t. [HMO.sub.it]:
Percentage of the total population enrolled in an [HMO.sub.it]
for state i in year t. [Uninsured.sub.it]: Percentage of the
total population uninsured for state i in year t. [Poverty.sub.it]
: Percentage of the total population in poverty for state i in
year t. [Labor.sub.it]: Labor force divided by civilian
noninstitutional population for state i in year t. [Unemp.sub.it]:
Unemployment rate for state i in year t. [Birthrate.sub.it],: Live
births per 1000 in population for state i in year t. [NME.sub.t]:
Number of new molecular entities approved by the FDA in year t.
NME[(-2).sub.t]: Number of new molecular entities
approved by the FDA in year t - 2.

Table 2. Fixed-Effects Regression Results
for Instruments: HMO and Uninsured

 HMO HMO

Variable 1 2

Obese 0.740 *** (0.160) 0.850 *** (0.162)
Smoke -0.204 (0.158) 0.123 (0.157)
Alcohol -3.144 (2.17) -4.100 * (2.17)
Income 2.298 ** (0.326) 2.683 *** (0.378)
Unemp. -0.472 * (0.269) -0.394 (0.277)
NME 0.083 *** (0.024) --
NME (-2) -- -0.047 (0.030)
Labor -1.151 *** (0.229) -1.159 *** (0.232)
Poverty -- --
Birthrate -- --
N 450 450
[R.sub.2]
 (group effects) 0.74 0.74
[R.sub.2]
 (X variables) 0.42 0.41
[R.sub.2]
 (X and group 0.90 0.9
 effects)
Adjusted
 [R.sub.2] 0.89 0.88
F value 63.18 *** 61.56 ***

 Uninsured Uninsured

Variable 3 4

Obese 0.010 (0.063) -0.011 (0.063)
Smoke 0.007 (0.062) 0.006 (0.062)
Alcohol -1.901 ** (0.869) -1.842 ** (0.859)
Income 0.576 *** (0.128) 0.456 *** (0.146)
Unemp. 0.164 (0.109) 0.136 (0.111)
NME -0.010 (0.010) --
NME (-2) -- 0.018 (0.012)
Labor --
Poverty 0.268 *** (0.047) 0.263 *** (0.047)
Birthrate -- --
N 450 450
[R.sub.2]
 (group effects) 0.83 0.83
[R.sub.2]
 (X variables) 0.48 0.49
[R.sub.2]
 (X and group 0.86 0.87
 effects)
Adjusted
 [R.sub.2] 0.85 0.85
F value 44.91 *** 45.09 ***

 65+ Years 65+ Years

Variable 5 6

Obese -0.010 (0.006) -0.013 ** (0.006)
Smoke -0.005 (0.006) -0.003 (0.006)
Alcohol -0.282 *** (0.101) -0.276 *** (0.099)
Income -0.059 *** (0.012) -0.083 *** (0.014)
Unemp. 0.011 0.005 (0.010)
NME 0.0001 --
NME (-2) -- 0.004 *** (0.001)
Labor -- --
Poverty -- --
Birthrate -0.224 *** (0.017) -0.229 *** (0.017)
N 450 450
[R.sub.2]
 (group effects) 0.99 0.99
[R.sub.2]
 (X variables) 0.44 0.44
[R.sub.2]
 (X and group 0.99 0.99
 effects)
Adjusted
 [R.sub.2] 0.99 0.99
F value 1419.73 *** 1466.21 ***

Dependent variables: [HMO.sub.it]: Percentage of the total
population enrolled in an HMO for state i in year t.
[Uninsured.sub.it]: Percentage of the total population
uninsured for state i in year t. 65+ [Years.sub.it]:
Percentage of the total population 65 years of age and
older for state i in year t. All cross-section estimates
are suppressed. Standard errors in parentheses.

* = significant at 0.10.

** = significant at 0.05.

*** = significant at 0.01.

Table 3. Fixed-Effects Regression Results for Real Per
Capita Prescription Drug Spending

 Fixed Fixed
 Effects Effects

Variable 1 2

Obese 3.67 *** (0.645) 3.18 *** (0.604)
Smoke -0.698 (0.633) -0.035 (0.582)
Alcohol -2.62 (9.87) -10.22 (9.23)
Income 24.27 *** (1.42) 18.94 *** (1.51)
Unemployed -2.09 ** (1.09) -3.37 *** (1.04)
HMO # 1.17 *** (0.198) 1.31 *** (0.184)
Uninsured # 0.990 ** (0.489) 0.655 (0.459)
65+ Years # 22.87 *** (4.51) 19.50 *** (4.24)
NME 0.107 (0.098) --
NME (-2) -- 0.847 *** (0.112)
N 450 450
[R.sup.2] (group
 effects only) 0.33 0.33
[R.sup.2] (X
 variables only) 0.74 0.78
[R.sup.2] (X and
 group effects) 0.93 0.94
Adjusted R2 0.92 0.93
F value 94.76 *** 109.27 ***

 IV Fixed IV Fixed
 Effects Effects

Variable 3 4

Obese 2.497 *** (0.908) 2.02 ** (0.891)
Smoke -0.142 (0.749) 0.294
Alcohol 9.11 (14.555) 9.46
Income 21.88 *** (2.51) 18.38 *** (2.55)
Unemployed -0.882 (1.44) -2.318 * (1.30)
HMO # 2.87 *** (0.894) 2.39 *** (0.820)
Uninsured # -0.169 -0.979
65+ Years # 29.80 *** (9.06) 36.48 *** (8.13)
NME -0.050 (0.133) --
NME (-2) -- 0.874 *** (0.133)
N 450 450
[R.sup.2] (group
 effects only) 0.33 0.33
[R.sup.2] (X
 variables only) 0.74 0.78
[R.sup.2] (X and
 group effects) 0.92 0.93
Adjusted R2 0.91 0.92
F value 76.37 *** 91.17 ***

 AR & IV Fixed
 Effects

Variable 5

Obese 1.63 * (0.856)
Smoke 0.632 (0.455)
Alcohol 9.93 (12.294)
Income 22.99 *** (2.39)
Unemployed -3.689 *** (1.13)
HMO # 0.189 (0.814)
Uninsured # -1.118 (1.35)
65+ Years # 21.86 ** (10.89)
NME --
NME (-2) 0.653 *** (0.098)
N 400
[R.sup.2] (group
 effects only) 0.33
[R.sup.2] (X
 variables only) 0.68
[R.sup.2] (X and
 group effects) 0.88
Adjusted R2 0.87
F value 45.19 ***

*** Dependent variable: [Drug.sub.it] = Real per
capita prescription drug expenditures (1996 dollars)
for state i in year t. All cross-section estimates are
suppressed. # = Instrument in Equations 3, 4, and 5.

Standard errors in parentheses.

* = significant at 0.1.

** = significant at 0.05.

*** = significant at 0.01.