The impact of policy uncertainty on the cost of capital
for companies in the U.S. healthcare industry
Author: Dennis van Dijk Student ID: 1890964 Supervisor: dr. J. Mierau MSc. Thesis Finance
University of Groningen
Faculty of Economics and Business
Abstract
This paper shows that the U.S. healthcare industry has no net exposure to shocks in policy uncertainty on an industry level. Hence, the large risk-adjusted returns on the U.S. healthcare portfolio are not explained by investors commanding a premium for exposure to this systematic risk factor. The results are obtained by incorporating a quantitative index for policy uncertainty in the standard asset pricing models, i.e. the CAPM, Fama-French three-factor and Carhart four-factor. Estimates from the asset pricing models further suggest that government intervention risk does not become more important in determining the cost of capital of R&D intense sub-industries, relative to the health services sub-industry. Applying our methodology to all five major industry portfolios suggests that most industries have a zero net sensitivity to shocks in policy uncertainty on this level of aggregation. Policy uncertainty does, therefore, not explain any differences in risk-adjusted returns between industries. The results do not support the prevailing belief that governments indirectly put a brake on medical innovation. Furthermore, for companies in the medical sector it is important to realize that investors do not seem to value government intervention risk, so neither should they in determining their cost of capital.
Keywords: Risk premia; Medical innovation; Government policy; Political uncertainty; Asset pricing JEL-classification: G12; G18,
1
1.
Introduction
In one of his first press conferences in January 2017, the then U.S. president-elect, Donald Trump stated that pharmaceutical companies are “getting away with murder” because of the generally high prices on drugs. The “big pharma” sector saw their stock prices fall directly after1. And it’s not the first time that the medical sector is under attack for their, sometimes outrageous, prices. Martin Shkreli, a former hedge fund manager, caused a heated social debate after raising the price of Daraprim, a 62-year old drug used by cancer and AIDS patient, by more than 5000%. One of the most striking facts about price setting in the industry is that the largest buyer of drugs in the world, the U.S. government, is forbidden by Federal law to use its bargaining power to bring down drug prices for beneficiaries in the Medicare and Medicaid programs.
However, from a business perspective, it is especially the profitability of monopoly pricing that makes the investments in new medical innovations worthwhile. Furthermore, academic research shows that the high medical costs might just be a logical consequence of our high standard of living nowadays. We become richer and richer over time and additional consumption does no longer raise our utility as much as spending money to live longer and enjoy our wealth longer. An important paper by Koijen, Philipson and Uhlig from 2016 even suggests that the constant threat of the government intervening in the market may even put a brake on large welfare gains for society. They state that investors in the healthcare industry currently require an addition 4%-6% on their equity stake to compensate for potential losses due to government intervention. This “medical innovation premium” would have the results that investments in medical innovation are half of what they would be under optimal conditions. In the best-case scenario, healthcare spending would be even higher by 4% than it is currently. This explanation by Koijen et al. can have far reaching consequences for society, policy makers and investors alike. However, their core assumption that especially the required returns on healthcare stocks are prone to policy-uncertainty, remains fairly unverified because uncertainty is difficult to quantify.
Recent advances in quantifying policy uncertainty by Baker, Bloom and Davis (2016) have created the opportunity for this research to revaluate the fundamental assumption underlying the described role of government intervention risk on the medical industry. This research sets out to answer the question whether compensation for increased exposure to policy uncertainty decreases the persistent excess returns on U.S. healthcare equity. In answering this research question, I first show that the aggregated portfolio of the healthcare industry shows no net exposure to policy uncertainty over the period 1986 to 2017. Furthermore, the large risk-adjusted returns prevail, even in the shorter sub-periods where the healthcare industry seems to behave like how Koijen et al. (2016) expects it to. I continue by showing that the zero net exposure to policy uncertainty is not caused by aggregating required returns on the R&D driven and the service oriented sub-industries. All portfolios on a sub-industry level convey the same message as the total healthcare industry portfolio, that there seems to be no net exposure to this particular risk.
In addition, I address the issue that investor compensation as a response to policy uncertainty may show larger differences in risk-adjusted excess returns. Only the consumer goods industry is shown to have a significant net exposed to this type of risk, causing no large changes in the perceived relative profitability of the different industries for investors, in terms of alpha.
1 Article on Reuters, among other,
2
Any concerns at this stage about the policy uncertainty factor not capturing policy related risk, are dismissed based on the finding that the employed risk factor is found to be priced in the cross-section of stock returns. For a set of 25 test assets that span the space of U.S. stock returns, I estimate a risk premium for policy uncertainty of 0.99% per month, resulting in a spread of 4.16% in annual required return as a consequence of exposure to policy uncertainty risk.
Although exposure to policy uncertainty risk does not explain the medical innovation premium, inclusion of this priced risk factor in consumption-based asset pricing models might enable investors to make better predictions about future stock returns in the healthcare industry. However, I show that the out-of-sample performance of the empirical Carhart four-factor model is not significantly improved when including the policy uncertainty factors.
The remainder of the paper is organised in the following way. Section 2 provides an overview of the major stakeholders in the healthcare industry and their incentives. This shows the importance of research to risk factors and cost of capital estimates in the medical industry and places this research in this context. Section 3 elaborates on the methodology and data. Section 4 presents the main results and provides a discussion of the findings. Last, section 5 concludes.
3
2.
Theoretical motivation
This chapter sets out to provide the context and theoretical motivation for this research. The first three sections show that the debate about rising healthcare expenditures revolves around the often conflicting interests of society, firms and investors. A recent paper with important implications for all three stakeholders is provided by Koijen et al. (2016). The authors suggest that the abnormally high historical returns on healthcare investments are a compensation for the large uncertainty about future government policies. If this premium were to be (partially) eliminated, opportunity cost for R&D investments would decrease, causing an increase in new profitable medical innovations. Ultimately, the advances in medical innovation would result in better and cheaper healthcare, with large gains for society. However, as will be shown in section 2.4, the evidence in favour of this explanation is highly suggestive and fails to estimate the actual impact of government intervention risk on healthcare returns, mainly because of the difficulty in quantifying policy uncertainty. Section 2.5 introduces the recent achievement by Baker et al. (2016) in quantifying policy uncertainty, which provides the opportunity to test the assumptions underlying the government intervention risk explanation. Given the large potential impact for different stakeholders, section 2.6 shows that it is important to formulate a quantitative answer to whether compensation for increased exposure to policy-related economic uncertainty decreases the persistent excess returns on U.S healthcare equity.
2.1 Rising healthcare spending
One of the major concerns in modern society is the continuous growth of public and private healthcare expenditures. The most recently published annual OECD health statistics of November 2017 show that people in OECD countries live longer (80.6 years, on average), but the burden of mental illness and chronic diseases is rising. In 2016, total health spending accounts for 9% of GDP on average, with extremes of 4.3% in Turkey and 17.2% in the United States (OECD, 2017). The average spending on health in the OECD was about $4000 per person, while the per capita expenditures in the U.S. were as high as $10,000. The increase over time of the health share for four different OECD countries is
Figure 1: Total expenditure on health as share of gross domestic product
The graph shows the increasing health share for Canada, Germany, the UK and the US over the time period 1970 to 2016. Health share is defined as the percentage of GDP flowing to total health expenditures. Source: OECD health statistics 2017. 0 2 4 6 8 10 12 14 16 18 1970 1975 1980 1985 1990 1995 2000 2005 2010 2015 2020 H ea lth s ha re Year United States Germany Canada United Kingdom
4
shown in figure 1. An aging population, rising incomes and advances in medical innovation are most often denoted as the key drivers behind this growth in health expenditures (Jenker and Leive, 2010).
According to the OECD data, the United States is accountable for the disproportionate amount of 49.1% of total healthcare spending in 2016, while it’s GDP equals just 24.6% of world GDP in the same year2. This concentration of world health care spending is a consequence of a free-rider problem introduced by the finding that medical innovation is driven predominantly by total world returns (Egan and Philipson, 2013). They argue that, with medical innovation being a public-good, small countries have nothing to gain from raising reimbursements as they will not significantly increase world returns to medical innovation and will nonetheless benefit from new innovations. However, an important implication of the world-profits on medical R&D being concentrated in the U.S, is that healthcare spending growth for any country depends on medical innovation and hence on U.S. policy.
The concentration of healthcare spending in the U.S and the resulting importance of U.S healthcare policy causes the majority of health economics research to focus on the U.S market. Further describing this market, the U.S. is the only OECD country where government schemes and compulsory health insurance are not the main source of healthcare financing. Reports by the Centers for Medicare and Medicaid Services (hereafter CMS) state that 20% of U.S. health spending was funded publicly through Medicare and 17% through Medicaid (CMS, 2017). Although most of reimbursements were done via Medicare and Medicaid, the households ultimately paid the largest share (28%) of total health spending, either through out-of-pocket spending, contributions to private health insurance premiums and contributions to Medicare through payroll taxes and payment of premiums (CMS, 2017; Hartman et al, 2018). The total healthcare spending increased with 4.3% in 2016, a deceleration of growth after faster growth rates in 2014 and 2015 of respectively 5.1% and 5.8%, associated with the coverage expansion under the Affordable Care Act (ACA) and fast increases in prescription drug spending (Hartman et al, 2018). Lastly, the three largest categories of medical services and goods are hospital care (32%), physician and clinical services (20%) and prescription drugs (10%). The growth within these three spending categories was mainly driven by medical price growth and increased intensity of health service and goods consumption. This situation, where the majority of reimbursement decisions are made by federal programs, who “lets” medical prices and intensity increase further, while the burden is carried mostly by households, forms the basis for public concern about the growing health share.
Now should the debate on health policy focus on limiting growth of health spending? Hall and Jones (2007) show that increased health spending is a direct consequence of growth in income. They argue that health spending is a superior good with an income elasticity above unity. Hence, marginal utility of life extension does not decline in wealthy societies, while marginal utility of consumption falls. The composition of total spending will therefore shift naturally towards health as people get richer. However, the prevalence of a third-party payer causes the demand for healthcare to be artificially price insensitive. This introduces different inefficiencies, such as overconsumption, higher equilibrium prices and a lower willingness to invest in cost-decreasing medical innovations (Buff and Terrell, 2014; Murphy and Topel,2006). Further evidence suggesting inefficiency in the healthcare markets comes from the persistent abnormally high profit margins in the pharmaceutical industry. In 2002, Angell already noted that the top 10 drug companies in the U.S. had a median profit margin of 17%, compared to 3.1% for all other industries on the Fortune 500 list. In 2016, the three largest pharmaceuticals by revenue were Johnson and Johnson ($72 bn), Roche ($53 bn) and Pfizer ($53 bn), with profit margins of respectively 23.0%, 18.2% and 13.7%3. The medical industry justifies this abnormal profitability by pointing at their larger dependency on innovation and the risks associated with it. Hence, cost-benefit analysis from both a societal and firm perspective might contribute to the healthcare debate by providing valuable insights in the incentives for innovation and inefficiencies in pricing and allocation.
2 GDP estimate obtained from World Bank: data.worldbank.org
5 2.2 Cost-benefit analysis of medical innovation
The allocation and the assessment of the adequacy of public healthcare spending is inherently complex. In addition, a government intervening in the marketplace to establish cost-effectiveness might result in a situation where drug manufacturers and other producers of medical technology lack sufficient incentives to innovate. In such a case, the policy effectively substitutes welfare of future patients who cannot be treated with the new blockbuster drugs that could have been developed, with the welfare of patients today who benefit from cheaper drugs.
Murphy and Topel (2006) develop a methodology to assess the value of medical research and health-augmenting innovations for society. They estimate both the value of raising the quality of life and decreasing mortality rates. As an example, they show that in the U.S. alone, a permanent 1 percent reduction in mortality from cancer can have a present value of $500 billion for society. Furthermore, they argue that in 2000, the U.S. medical expenditures since 1970, totalling to an amount of $34 trillion, yielded a gross social value of $95 trillion. These large gains in welfare justify substantially greater investments than are actually observed. Moreover, this “missing R&D puzzle” suggests that there is a disparity between the net present value accruing to the firm and the gain in welfare for society.
The R&D cycle in the pharmaceutical, biotechnology and medical device sectors, can be characterized by high initial costs, long development times, low success rates per project and, hence, a large dependency of profitability on the cost of capital. Within the pharmaceutical industry, leading research by DiMasi et al. (2003) finds an out-of-pocket cost per approved drug of $403 million during the period 1983 to 1994. In addition, they estimate a success rate of 21.5% of drugs brought into clinical trials. The typical time span for the clinical trials, from the “Investigation of New Drug permit” (IND) to “New Drug Approval” (NDA), may take up to seven years with an additional two years for regulatory approval (Scherer, 2010). After approval, a new drug is not by definition a commercial success. Grabowski and Vernon (1990, 1994) have shown that the distribution of profitability is highly skewed, with 20% of the most profitable new drugs responsible for over 70% of total generated present value. Moreover, only the top three deciles had present values in excess of average R&D costs. The mean real IRR for R&D projects during the 1980’s, as analysed by Grabowski and Vernon (1994), was estimated to be 11.1%. Comparing to a real cost of capital estimate of 10.5%, yields a typical NPV of $22.2 million (in 1990 dollars). However, sensitivity analysis shows that the net value of a typical R&D project quickly erodes when the contribution margin falls below 40% or when outside investors require a slightly higher return on equity.
2.3 The importance of reliable cost of capital estimates
The result of Grabowski and Vernon (1994) suggests that pharmaceutical research has a net present value close to zero. As a consequence, realistic estimates of the cost of capital and an understanding of which factors influence this opportunity cost, are immensely important for project development and regulatory policy decisions. Myers and Shyam-Sunder (1996) show that the cost of capital used by a firm to make R&D investment decisions mainly depends on the required return on equity, because debt ratios are exceptionally low in the sectors concerned. A comprehensive overview of cost of capital estimates and systematic risk factors from historical healthcare stock returns, is given by Harrington and Miller (2009).
The overview given by Harrington and Miller (2009) shows that the required equity returns are mostly obtained using the Sharpe (1964) CAPM and Fama and French (1992) three-factor asset pricing models. Both models are based on the notion that investors require a compensation in the form of a risk premium for holding a security with a certain exposure to an undiversifiable risk. The CAPM measures the sensitivity (expressed as beta) of the firm’s return relative to the return on the market
6
portfolio of risky assets. Hence, the market beta is a measure of the security’s exposure to the undiversifiable market risk. In addition, the Fama French model also estimates betas with respect to a size factor and book-to-market factor. Both the size factor and the book-to-market factor are empirical risk factors based on the observation that stock returns of small companies outperform those of large companies and value stocks (high B/M) outperforms growth stocks (low B/M).
Estimates of market betas for healthcare firms are on average slightly below or around one. The Office for Technology Assessment (1993) reports a beta of 0.90 for pharmaceuticals during 1975-1987. The cost of capital estimates of Myers and Shyam-Sunder (1996) for large pharmaceuticals are based on CAPM betas of 0.98, 0.70 and 1.04 during the five-year periods starting in 1975, 1981 and 1985, respectively. Golec and Vernon (2007) estimated the Fama-French model with stock market data of pharmaceutical and biotechnology firms during the period 1982-2005, and found average market betas of 0.92 for pharmaceuticals and 1.06 for biotech. In addition, Golec and Vernon (2007) further show that returns on both pharmaceuticals and biotech stocks have a large and significant loading on the size factor. This size-related risk is especially pronounced in the, on average, smaller biotech firms and results in a higher cost of capital for these firms. Hence, the equilibrium level of R&D spending in biotech is expected to be lower than in well-established pharmaceutical companies.
Based on monthly return series between 2001 and 2008 of pharmaceuticals, biotech and medical device firms, Harrington and Miller (2009) argue that, although, the scientific uncertainty of R&D projects is diversifiable for investors, the R&D investments have a systematic market-risk component that is not. They proceed by showing that R&D intensity is positively related to market betas, after controlling for sector. Another important finding is that the inter-sectoral differences in average market betas and exposure to other systematic risk factors, after controlling for R&D intensity, are significant and change over time. For example, biotech firms had significantly higher market betas than pharmaceutical and medical device firms during the 2001-2005 period, while medical device firms had significantly higher betas than pharmaceutical firms within the 2006-2008 window (Harrington and Miller, 2009).
The scope of the studies mentioned above is to quantify uncertainty in the cost of capital estimates. However, most of these studies neglect the significant overcompensation of investors for the risk exposures in their respective models. An analysis of the U.S. healthcare industry as an investment opportunity is provided by Tresl et al. (2014). They find that the healthcare industry portfolio earns the highest factor risk-adjusted return (expressed as alpha) of 1.8% - 3.5% annually. Moreover, this result became more pronounced since the mid 1980’s. At the same point in time, medical inflation began to outpace the general CPI and healthcare returns became less correlated to market returns. Based on the measured risk exposures, they argue that the healthcare industry post 1985 can be characterized as slightly larger, less mature and with a positive exposure to the momentum characteristic. Placed in the broader context, the result by Tresl et al. (2014) suggests that investors either require compensation for an undefined systematic risk or that investors persistently under-price healthcare companies.
2.4 The medical innovation premium and government intervention risk
To summarize the discussed literature, the debate about healthcare spending revolves around three interrelated questions:
1) What is the optimal amount of healthcare spending? 2) Do we invest enough in medical R&D?
7
An important new insight that encompasses all facets of the healthcare debate is given by Koijen et al. (2016). They argue that the excess health equity return of 4-6% annually is a premium for exposure to government intervention risk. Without this government-induced profit risk, the R&D spending share (as % of GDP) would be twice as high and health spending would be higher by 4% of GDP. The long-run equilibrium health spending share in the absence of government intervention risk would be 38%. The risk-adjusted excess return of 4-6% on healthcare portfolios, estimated using CAPM and Fama-French, is dubbed the “medical innovation premium” because the highest premiums were found for the R&D intense pharmaceutical and medical devices sub-industries.
Koijen et al. (2016) considers government intervention risk a plausible explanation for the medical innovation premium because of three reasons. First, government policy sets FDA approval requirements and reimbursements through Medicare and Medicaid. Second, government intervention risk is an aggregate risk component to which the healthcare sector is particularly exposed. Third, other plausible risk factors, such healthcare productivity and longevity, will generate a negative alpha in consumption-based asset pricing models.
Empirical evidence in support of the government intervention risk explanation consists of the following three findings:
1) Analysis of the “risk factor” section in the 10-K filings of 50 health and non-health companies shows that healthcare firms mention non-healthcare specific government related words more often. 1.51% of total word count versus 1.23%, on average.
2) Three large drawdowns of the health care sector are identified and these periods coincide with the discussion of the Clinton health care reform (1992-1993), the technology crash (2000-2002) and the financial crisis (2007-2008).
3) An event-study approach shows average cumulative abnormal returns for healthcare stocks of -23.6%, following key events around Clinton’s health care reform.
The fundamental problem underlying the work by Koijen et al. (2016) is the difficulty of quantifying government intervention risk. Without a proxy for this risk factor they are unable to verify explicitly the underlying assumptions that the healthcare sector is particularly exposed to this risk and that exposure to this risk factor generates a positive premium in consumption-based asset pricing models that decreases the alphas. Furthermore, their quantitative predictions are based on a long-run general equilibrium model of household consumption, entrepreneurial activity and government subsidies. However, alpha enters as an input and a fixed probability of government intervention is “artfully” assumed in the order 10%-20%. The model is calibrated to the historical healthcare share and R&D share using four additional fitting parameters such as initial level of medical knowledge and curvature of the R&D production function. The model provides valuable insights in potential mechanisms through which government intervention risk affects healthcare spending and R&D investments, but should not be mistaken for conclusive evidence that alpha is fully explained by compensation for government intervention risk.
2.5 Quantifying economic policy uncertainty
Baker, Bloom and Davis (2016) are the first to construct a measure that allows continuous tracking of policy risk. They created an index of policy-related economic uncertainty based on newspaper coverage frequency. Moreover, they show that increases in their policy uncertainty factor are associated with greater stock price volatility, and decreases in industrial production, employment, GDP and real investment. The economic policy uncertainty (EPU) index can be decomposed into category specific uncertainty, such as national security, healthcare and monetary policy. The largest
8
sources of elevated EPU in recent years are fiscal policy (taxes and government spending) and healthcare policy.
Pástor and Veronesi (2013) and Brogaard and Detzel (2015) also acknowledge that government policy is ubiquitous and, hence, most likely undiversifiable. Therefore, they both test the asset-pricing implications of the Baker et al. EPU factor. Pástor and Veronesi (2013) show that economic policy uncertainty is counter-cyclical and they, therefore, test whether exposure to EPU commands a risk premium conditional on the state of the economy. They find a marginally significant relation between 3-,6- and 12-month future excess market returns and the EPU index during weak economic conditions. Brogaard and Detzel (2015) focus on the unconditional relationship between EPU and expected returns on the U.S market portfolio. They show that innovations in economic policy uncertainty command a significant negative risk premium in the cross-section of returns. The spread in EPU sensitivity within a set of test assets accounts for a difference in excess return of 5.53% per annum.
The literature shows that the newspaper-based index of Baker et al. (2016) does capture policy uncertainty to some extent and that exposure to this risk factor commands a significant premium in the overall U.S. stock market. This motivates further research into industry-specific exposure to EPU and allows testing of the fundamental assumptions of Koijen et al. (2016) about the healthcare industry.
2.6 Research outline
This research sets out to test the government induced profit risk explanation for the medical innovation premium presented by Koijen et al. (2016), by employing the recently developed quantitative measure of policy uncertainty by Baker et al. (2016). This results in the following research question:
“Does compensation for a relatively large exposure to policy-related uncertainty explain the persistent risk-adjusted excess returns on U.S healthcare equity?”
To avoid ambiguity, I will elaborate on the interpretations and definitions of the following key terms from the research question:
“Compensation” is the premium required by investors for holding a security with a larger perceived risk.
“Relatively large” refers to a direct comparison with the other four major industries. Koijen et al. (2016) expect that the healthcare industry is the most sensitive, resulting in the largest risk-adjusted excess returns for all five industry portfolios.
“Risk-adjusted excess return” and “exposure” are defined as the intercept (or alpha) and betas with respect to the risk factors in an asset pricing model.
“Policy-related uncertainty” is the risk factor derived from the Baker et al. (2016) Economic Policy Uncertainty (EPU) index by taking the unexpected component.
“U.S. healthcare equity” is one of the five major industry portfolios, following the widely used SIC-code based definition of Kenneth French.
An answer to this research question shows policymakers to what extent their interference with healthcare markets raises the cost of capital faced by firms in this sector and, consequently, decreases further investments in medical innovation. Any effort in lowering the cost of capital for medical innovation can result in large welfare gains for society.
9
This research contributes to the existing academic literature in two ways. First, it extends the literature about the interaction between policy uncertainty and stock markets by examining differences in exposure on an industry level. Second, it adds to the literature that analyses cost of capital estimates and risk exposure for investments in medical R&D, by comparing the performance of different consumption-based asset pricing models.
The relevance for investors is twofold. First, this study provides insights in the ex-post performance of the healthcare portfolio and the size of the risk-adjusted excess returns that prevail after accounting for exposure to government intervention risk. Second, the increased forecastability of healthcare returns is assessed.
10
3.
Data and Methodology
This research sets out to examine the medical innovation premium and its proposed origin in exposure to policy-related uncertainty. In order to provide a comprehensive answer to the main research question, section 3.1 will derive multiple smaller sub-questions. Next, the methodology will be described in section 3.2. To ensure comparability with earlier work, I will closely follow the methodology used by Koijen et al. (2016) to quantify the premium. Last, section 3.3 provides an overview of the data and shows how the uncertainty variables are constructed.
3.1 Methodology
The purpose of this research is to provide an answer to whether compensation for a relatively large exposure to policy-related uncertainty explains the persistent risk-adjusted excess returns on U.S healthcare equity. An all-encompassing answer can be established after answering the smaller, more specific sub-questions shown below. The remainder of this report will follow the structure specified by the sub-questions.
Section 4.1:
Has the healthcare industry portfolio a significant beta with respect to the policy-uncertainty factor?
Does the asset pricing model still show a significant risk-adjusted excess return for the healthcare industry portfolio after accounting for policy uncertainty?
Section 4.2:
Is exposure to the policy uncertainty risk factor related to medical innovation and, hence, more pronounced in the R&D intense sub-industries?
Section 4.3:
Does an asset pricing model with the policy uncertainty risk factor decrease the industry differences in risk-adjusted excess return?
Section 4.4:
Is policy uncertainty priced in the cross-section of U.S equity returns when considering a set of test assets that effectively span the mean-variance frontier of the U.S. stock market?
Section 4.5:
Is the in-sample accuracy of the asset pricing model with uncertainty factors verified out-of-sample by its ability to accurately forecast healthcare stock returns?
3.2 Asset pricing theory
Koijen et al. (2016) reports the medical innovation premium on a risk-adjusted basis, as the part of the time-series average expected excess return that is not explained by two common benchmark asset pricing models (given by the intercept, or alpha). The asset pricing models employed are the Sharpe (1964) Capital Asset Pricing Model (hereafter: CAPM) and the Fama and French (1992) three-factor model (hereafter: FF3).
3.2.1 – General Asset pricing theory
Following the standard asset pricing literature (e.g. Cochrane, 2009), any conditional asset pricing model can be expressed as:
11
𝐸 [𝑚 𝑟, ] = 0 (3.1)
With Et[.] the expected value operator conditional on information available at time t, mt+1 the stochastic discount factor (SDF), that can also be interpreted as an investors marginal rate of substitution, and ri,t+1 the excess return on an asset i. Assuming that the “law of one price” and the concept of “absence of arbitrage” hold, it can be stated that there is a discount factor that prices all assets and that this discount factor has to be positive.
Linear factor models, including the CAPM, ICAPM and FF3, additionally specify a stochastic discount factor of the form:
𝑚 = 𝑎 + 𝑏 (𝑓 − 𝐸 [𝑓]) (3.2)
Where b’ denotes a vector of coefficients, a is a constant and ft is a vector of factors that act as proxies for marginal utility growth. To be economically sensible, mt should be high during recessions or “bad times” and low during “good times”. Furthermore, assets that covary counter-cyclically with the SDF provide higher pay-offs during times that are already good for investors and lower pay-offs in bad states of the economy. A lower price for these less desirable assets increases their expected returns, effectively resulting in investors requiring a premium for holding these assets that covary negatively with marginal utility. A similar line of reasoning is valid for assets that covary positively with the SDF, for which investors are willing to pay the premium to hedge.
The constant, a, in equation (3.2) can be normalized arbitrarily if the assets are expressed as excess returns. If normalized to a=1, then one can find the price of risk factors, λ, such that:
𝐸 𝑟, = 𝛽 𝜆 (3.3)
or;
𝐸 𝑟, =
, ,
( ) −𝑏 𝑣𝑎𝑟(𝑓 ) (3.4)
Hence, an SDF that is linear in its risk factors is equivalent to the expected return-beta representation of linear multi-factor models that is more common in the empirical asset pricing literature. The beta terms of equation (3.3) are a measure of contemporaneous exposure of an asset i’s excess return to each factor, defined as the coefficients in a time series regression of excess returns on factors:
𝑟, = 𝑎 + 𝛽 𝑓 + ϵ (3.5)
The betas subsequently explain the variation in average excess returns across assets and allow estimation of the factor risk premia from a cross-sectional regression of time-averaged excess returns on betas:
𝐸 (𝑟 ) = 𝛽 λ + α (3.6)
From equation 3.3 it follows that the pricing error α of the cross-sectional regression in equation (3.6) should be zero. The intercept of the time-series regression 𝑎 is not necessary equal to the pricing error α.
Important for the inclusion of the uncertainty factors is the notion that, If the factor 𝑓 is expressed as an excess return series, then this asset will have a beta of one with itself and from equation (3.6) it follows that the price of risk will be equal to the expected value of the risk factor:
12
𝜆 = 𝐸 (𝑓) (3.7)
Only if the factor is expressed as an excess return, then the 𝑎 of the time-series regression should go to zero and can be used to form the Gibbons Ross and Shanken (GRS) test statistic that evaluates the asset pricing model by testing if the time-series intercepts of all test assets are jointly zero. However, when estimating the price of the risk factor λ, the two-stage regression provides a more robust estimate as it allows non-zero pricing errors for factors and minimizes the sum of squares of all pricing errors.
3.2.2 – Implications of literature on methodology
The above analysis shows that it is imperative to create a factor mimicking portfolio when including a non-return series as independent variable. The mimicking portfolio captures the policy uncertainty factor but is expressed in excess returns. This allows running time-series regression with policy uncertainty factors in a similar fashion as was done by Koijen et al. (2016). Taking this into account, the time-series alphas that are adjusted for policy uncertainty exposure, presented later in the results, can be compared with those provided by Koijen et al.
The candidate policy-uncertain factors are incorporated in the asset pricing models using the ICAPM framework by Merton (1973). Appendix A elaborates on the correct implementation of the uncertainty variables and tests whether EPU,VXO and HPU are relevant state variables. Furthermore, appendix C shows how the factor mimicking portfolios are created. These appendices provide a necessary justification of using the uncertainty variables, but are not directly relevant in answering the research question. Important characteristics of the factor mimicking portfolios and other results from the appendices that are necessary in answering the research question, will be summarized in the main body of the text.
Time-series regressions of the asset pricing model described in equation (3.5) are used to obtain estimates of the factor loadings and the risk-adjusted alphas. The different asset pricing models that are compared to each other are the popular Sharpe (1964) CAPM and the more robust Fama and French (1993) three-factor model and Carhart (1997) four-factor model.
In addition to the time-series estimates, I will also show that the policy uncertainty factor is a priced risk factor by estimating λ’s from a two-stage cross-sectional regression described by equations (3.5) and (3.6). These results are presented in appendix D and summarized in section 4.4.
3.3 Data and variable construction 3.3.1 – The Healthcare industry portfolios
In accordance to Koijen et al. (2016), I follow the Fama and French five-industry classification that splits the universe of U.S equity into: consumer goods, manufacturing, high-tech, healthcare and “other”, based on their respective standard industry classification (SIC) codes. In addition, following the Fama and French 48-industry classification, the healthcare industry includes three different sub-industries: health services, medical equipment and drugs. The monthly return data for the value-weighted (sub-)industry portfolios is obtained from Kenneth French’s homepage4. Lastly, an important benchmark for describing the healthcare portfolios is the general US stock market, for which I will use the value-weighted return of all US-based firms included in the CRSP database. This return series is constructed by Kenneth French. Definitions of the portfolios used are also provided in table 1.
13
Table 1: US market and industry portfolio definitions
Overview of the portfolios used throughout this thesis. Definitions are based on the Fama and French 5-industry and 48-industry classifications, as available on Kenneth French’s homepage.
Name Label Composition
US market Mkt Value-weighted return of all CRSP firms incorporated in the US and listed on the NYSE, AMEX, or NASDAQ (Fama and French, 1993).
FF 5-industry classification:
Healthcare Hlth Value-weighted return of all CRSP firms incorporated in the US with SIC codes: 2830-2839, 3693,3840-3859, 8000-8099
FF 48-industry classification:
Health services HlthServ,Hlth* Value-weighted return of all CRSP firms incorporated in the US with SIC codes:
8000-8099
Medical equipment MedEq Value-weighted return of all CRSP firms incorporated in the US with SIC codes: 3693, 3840-3849, 3850-3851
Pharmaceutical products Drugs Value-weighted return of all CRSP firms incorporated in the US with SIC codes: 2830-2830, 2831, 2833, 2834, 2835, 2836
*Health services is often labelled Hlth in the 48-industry classification. In order to avoid confusion between the 5-industry and 48-industry Hlth portfolios, I will denote the 48-industry health services portfolio with HlthServ.
Table 2: Descriptive statistics of the US market and industry portfolios
Descriptive statistics for the monthly excess return series of the five primary research portfolios. Returns are simple returns expressed in %. Excess returns are calculated using the 1-month treasury bill rate from Ibbotson and Associates as the risk-free rate. Panel A reports descriptive statistics over the sample period used in Koijen et al. (2016) and Panel B reports the same statistics for the same portfolios but in the time window used in the remainder of this report. The table shows the monthly arithmetic mean, monthly standard deviation, annualized Sharpe ratio and the minimum, median, maximum, skewness and excess kurtosis of the distribution.
Panel A: from 1961:01 to 2012:12 (Koijen et al.)
Mean Standard deviation
Sharpe ratio
Minimum Median Maximum Skewness Excess kurtosis US market (Mkt) 0.47 4.50 0.36 -23.24 0.81 16.10 -0.50 1.80 US healthcare (Hlth) 0.61 4.98 0.43 -21.06 0.67 29.01 0.04 2.39 - Health services (HlthServ)* 0.59 8.47 0.24 -39.61 0.68 35.89 -0.06 2.37
- Medical equipment (MedEq) 0.67 5.43 0.43 -21.16 0.82 20.52 -0.28 1.11 - Pharmaceuticals (Drugs) 0.63 5.09 0.43 -19.71 0.70 31.29 0.15 2.62
panel B: from 1986:02 to 2017:08
Mean Standard
deviation Sharpe ratio Minimum Median Maximum Skewness Excess kurtosis US Market (Mkt) 0.67 4.41 0.52 -23.24 1.16 12.47 -0.92 2.83 US Healthcare (Hlth) 0.86 4.63 0.65 -21.06 1.06 16.09 -0.38 1.43 - Health services (HlthServ)* 0.57 6.46 0.31 -32.03 1.09 20.66 -0.54 2.20
- Medical equipment (MedEq) 0.88 5.03 0.61 -21.16 1.01 15.95 -0.62 1.97 - Pharmaceuticals (Drugs) 0.91 4.85 0.65 -19.71 1.15 15.89 -0.25 0.93
14
Table 2 describes the distribution of the monthly excess returns on the portfolios under consideration, without any adjustments for systematic risk. Returns are in excess of the risk-free rate, as proxied by the 1-month Treasury bill rate obtained from Ibbotson and Associates, inc. Furthermore, Panel A of table 2 shows descriptive statistics for the sample period from Koijen et al., while panel B shows descriptive statistics for the more recent time window used in this study. This period is the result of data availability of the policy uncertainty factors.
Although average returns are slightly higher and standard deviations are lower for the more recent period, both panels do convey similar information about the distribution of excess returns. Comparing the healthcare industry with the market in panel B, shows that the healthcare portfolio earns a higher excess return, 10.4% per annum versus 8.0% per annum respectively. The volatility of both portfolios, however, does not differ much: 16.0% per annum versus 15.3% per annum. This results in a higher return per unit of risk, as measured by the Sharpe ratio: 0.65 for the healthcare portfolio and 0.52 for the market portfolio. When looking at the sub-industry portfolio characteristics, it can be seen that the mean-variance efficiency of the healthcare portfolio originates from the medical equipment and pharmaceutical products portfolios, while the health services portfolio underperforms.
Table 3: Alphas relative to CAPM and FF3
The table reports OLS estimates of two asset pricing models specified by:
𝑟, = 𝑎 + 𝛽 𝑓 + ϵ
The factors 𝑓 are the excess return on the CRSP value-weighted market portfolio (Mkt) in case of the CAPM and the market excess return (Mkt), return on the zero-cost “small minus big” factor (SMB) and the return on the zero-cost “high minus low” book-to-market factor (HML) in case of the FF3. The samples are from 1961:01 to 2012:12 (panel A) and from 1986:02 to 2017:08 (panel B). Portfolio returns are monthly and in excess of the risk-free rate. Alphas are reported as annualized percentages. The t-statistics are reported in parentheses and calculated using Newey-West (1987) adjusted HAC robust covariance matrices, with automatic lag length selection following Andrews (1991).
Panel A: from 1961:01 to 2012:12 (Koijen et al.)
Hlth HlthServ MedEq Drugs 𝛼 (annual, %) 2.55 4.74*** 0.91 -0.31 2.89 4.20** 2.97* 5.42*** (1.64) (3.39) (0.21) (-0.08) (1.51) (2.32) (1.78) (3.62) 𝛽 0.85*** 0.85*** 1.14*** 1.05*** 0.91*** 0.86*** 0.82*** 0.84*** (18.26) (20.78) (12.06) (11.53) (21.30) (21.09) (16.36) (18.24) 𝛽 -0.25*** 0.57** 0.06 -0.34*** (-4.65) (2.15) (1.05) (-5.42) 𝛽 -0.31*** 0.13 -0.24** -0.34*** (-3.07) (0.61) (-2.17) (-3.31) No. monthly obs. 624 624 522 522 624 624 624 624 adj. R2 0.59 0.63 0.39 0.43 0.57 0.59 0.53 0.59
panel B: from 1986:02 to 2017:08
Hlth HlthServ MedEq Drugs α (annual, %) 4.22** 4.85*** 0.15 -1.32 3.77* 4.14** 4.94** 5.76*** (2.17) (2.73) (0.04) (-0.41) (1.81) (2.08) (2.27) (2.93) 𝛽 0.77*** 0.78*** 0.84*** 0.87*** 0.85*** 0.82*** 0.75*** 0.77*** (13.94) (15.40) (8.62) (9.51) (14.83) (15.39) (12.91) (13.93) 𝛽 -0.25*** 0.25 0.10* -0.36*** (-4.36) (1.16) (1.69) (-4.87) 𝛽 -0.21* 0.44 -0.09 -0.28** (-1.88) (0.42) (-0.67) (-2.46) No. monthly obs. 379 379 379 379 379 379 379 379 adj. R2 0.54 0.57 0.33 0.37 0.56 0.56 0.47 0.53 *** Denotes significance on the 1% level, ** on the 5% level and * on the 10% level.
15
Both sub-industries have in common that they heavily rely on medical R&D. Lastly, table 2 shows that all return series in panel B are slightly skewed-left and moderately leptokurtic. However, these minor deviations from normality provide no reason to discard the Sharpe ratio. Furthermore, central limit theorem suggests that the sample size (n=379 monthly returns) is sufficiently large for the mean to be normally distributed (Hamburg, 1970).
3.3.2 – Impact of different evaluation period on alphas
Koijen et al. (2016) reports the medical innovation premium on a risk-adjusted basis over the period 1961:01 to 2012:12. The period analysed in this research runs from 1986:02 to 2017:08. As the medical innovation premium forms the starting point of this study, I will already document the preliminary results of both time-series regressions here in table 3. This table shows annual alphas and factor betas with respect to the CAPM and FF3 model, estimated from monthly time-series. In contrast to Koijen et al., I estimate the models with monthly return data and report estimates of alphas together with the coefficients on the other factors to provide insights into the risk profile of each portfolio.
The OLS estimates of alpha for the healthcare industry from panel A (CAPM: 2.55%, FF3: 4.74%) are in good agreement with those reported by Koijen et al. (CAPM: 3.31%, FF3: 5.01%). Furthermore, panel B shows that the medical innovation premium is also significant and similar in size in the smaller, more recent time window (CAPM: 4.22%, FF3: 4.85%). Estimated coefficients with respect to the market portfolio are significant on the 1% level and smaller than 1, suggesting that healthcare portfolio is less volatile than the market. Furthermore, panel B shows a negative, significant exposure of the healthcare portfolio to the size and book-to-market factors. Again, the high excess return relative to the low risk exposures is reflected only in the medical equipment and pharmaceuticals sub-industry portfolios.
3.2.3 – Uncertainty factors
The goal of this research is to assess the medical innovation premium while controlling for any compensation related to policy uncertianty risk. To quantify policy uncertainty I use the recently developed economic policy uncertainty factor from Baker et al. (2016). Additionally, I evaluate the pricing implication of the categorial healthcare policy uncertainty index. Both the EPU and HPU factor are obtained from the Baker, Bloom and Davis homepage5. The raw data obtained has a monthly frequency and starts from 1985:01. The industry specific HPU factor is included in further analysis because Baker et al. (2016) showed that the categorial Healthcare Policy Uncertainty (HPU) has additional explanatory power in a contemporaneous regression of 30-day implied healthcare stock price volatility on EPU and HPU.
The literature review showed that Koijen et al. (2016) expects government intervention risk to be the aggregated risk factor underlying the medical innovation premium, partially based on an analysis of government-related word counts in healthcare 10-K filings. Likewise, the Baker et al. (2016) uncertainty index is also based on relative word counts. However, a comparison of both measures is impossible because of the difference in construction. The Baker et al. index is largely created by taking the ratio of news articles discussing policy uncertainty to the total volume of published articles in 10 large newspapers. In contrast, the word-count by Koijen et al. was based on annual 10-K filings for fiscal year 2006 and 2012. An uncertainty index based on 10-K filings would have the inherent drawback that it is restricted to an annual frequency.
Furthermore, Brogaard and Dretzel (2015) have shown that the effect of policy uncertainty on equity returns becomes more pronounced when controlling for general economic uncertainty with the
16
S&P 100 volatility index (VXO). This index is obtained from the Chicago Board Options Exchange6 and is available with a daily frequency, starting from 1986. The daily index is converted to a monthly frequency by taking month-end values. This reflects the timing of returns on the different portfolios, also commonly based on month-end prices.
The standardized explanatory variables EPU, HPU and VXO are plotted in figure 2. The graphs in figure 2 combined with correlation from table 4, shows that, although, all three uncertainty factors are positively correlated with respect to each other, distinct periods of elevated uncertainty can be identified within each factor. For example, the HPU factor shows clearly that there was increased healthcare policy uncertainty during the Clinton healthcare reform (1992-1994), during the implementation of the affordable care act (2007-2015) and following the Nov. 2016 elections. These periods are not necessarily coupled to periods of prolonged increased volatility on the stock market. In addition, the EPU index clearly shows additional periods of uncertainty that are not reflected in the HPU factor. For example, the uncertainty following black-Monday in Oct. 1987, the US subprime mortgage crisis (2007-2010) and additionally the peaks after the 9/11 terrorist attacks (2001) and the first and second Gulf war (1990-1990 and 2003, respectively).
As was stated in the methodology section, the EPU, HPU and VXO index are not expressed as returns. Therefore, appendix C describes how factor-mimicking portfolios are created for the EPU and VXO index. These factor mimicking portfolios capture the unexpected shocks in EPU and VXO, because the ICAPM literature states that only the unexpected component of a “state variable” requires a risk premium. The unexpected shocks, or “innovations”, in EPU and VXO are based on the residuals from an autoregressive model, following Campbell (1996) and Petkova (2006). The factor mimicking portfolio carries all the pricing information of the original factor (Cochrane, 2009) and reduces the possibility of spurious results.
Zero cost factor-mimicking portfolios have only been created for innovations in EPU and VXO, because HPU was found to have no additional explanatory power after including EPU and VXO. This conclusion is based on a regression of industrial production growth on the candidate uncertainty state variables, presented in appendix A. Moreover, Merton’s (1973) ICAPM framework suggests that a relevant state variable should be able to predict the aggregate return on wealth, for which I use the growth in macroeconomic activity as proxied by the US industrial production index from the Federal Reserve Bank of St. Louis (Cochrane, 2009). The EPU index is found to predict industrial production growth with a significant negative coefficient up to three months in the future. In addition, HPU does not provide additional explanatory power beyond the information share with EPU. As a robustness check, appendix B evaluates the forecastability of buy-and-hold returns instead of macroeconomic activity, similar as in Brogaard and Dretzel (2015). The results related to factor construction and selection are found to be robust under the different specification.
6 Daily VXO data from:
17
Figure 2: Time-series of standardized EPU, HPU and VXO
Graphs of the monthly time-series of Economic Policy Uncertainty (EPU), Healthcare Policy Uncertainty (HPU) and economic distress as measured by the month-end S&P 100 volatility (VXO), on the time window 1986:02 to 2017:08. The time-series are standardized to have zero mean and unit standard deviation.
a
rb
.u
n
it
Table 4: Correlation table EPU, HPU and VXO
This table reports Pearson correlation coefficients for the explanatory uncertainty variables: Economic Policy Uncertainty (EPU), Healthcare Policy Uncertainty (HPU) and month-end S&P 100 volatility (VXO). Correlation coefficients are applicable to the 1986:02 to 2017:08 time window and are estimated using monthly data.
EPU HPU VXO EPU 1 0.614*** 0.436***
HPU 1 0.078
VXO 1
18
4.
Results and discussion
The results will be discussed in the order of importance. As stated in the methodology, the first section starts by examining the static full-window exposure of the healthcare industry to uncertainty risk factors EPU and VXO. Next, as earlier research stated that risk exposures may have changed significantly over the period, the 36-month rolling-window regression results are subsequently evaluated. Section 4.2 decomposes the static and time-varying alphas and risk exposures per sub-industry. Section 4.3 compares the alphas and risk exposure of the healthcare industry with the other four major industries. Section 4.4 justifies inclusion of the uncertainty factors by showing that both factors command a risk premium in the cross-section of U.S. stock returns. The last section evaluates the incremental improvement in forecastability of healthcare returns using the uncertainty risk factors. 4.1 Risk exposure in the healthcare industry
4.1.1 – Static estimates for the period 1986 to 2017
Table 5 shows estimates of the CAPM, Fama-french’s three-factor model (FF3) and Carhart’s four-factor model, with and without the additional uncertainty variables. The dependent variable in the time-series regression is the monthly excess return on the healthcare industry portfolio (Hlth), following French’s five-industry classification.
For the full 1986-2017 window, annual healthcare alphas are in the range of 3.81% to 4.85%. The alpha estimates are significant on at least the 5% level. Accounting for the possibility that the observed risk-adjusted returns are simply a compensation for a size, value or momentum strategy, does not significantly alter the alphas. In addition, adding the uncertainty risk factors EPU and VXO to any of the standard asset pricing models, results in no significant reduction of alpha. After adding the uncertainty risk factors, the alpha with respect to the CAPM model slightly increases (+0.09%), while the alphas with respect to the more robust Fama-French 0.23%) and Carhart four-factor model (-0.18%) show a slight decrease. Hence, even after adjusting for policy uncertainty and general economic uncertainty, did the healthcare portfolio earn a statistically significant and economically meaningful ex-post risk-adjusted excess return.
Tresl et al. (2014) found an average rolling 36-month CAPM alpha for the healthcare industry over the period 1985-2010 of 2.1% per annum. Which is lower than the static 4.2% documented here. Further evidence in favour of time-varying alphas, comes from a comparison with the near-zero alpha estimates reported by Harrington and Miller over the periods 2001-2005 and 2006-2008. Both comparisons provide a first indication that market betas and excess returns have changed significantly over the sample period. This is indeed the case, as will be shown in section 4.1.2. However, the static alphas from table 5 are of the same order as the long-term (1961-2012) alphas reported by Koijen et al. (2016). This shows that the healthcare portfolio earns large significant risk-adjusted excess returns in the long-run, but holding periods shorter than four years can give different results.
Evaluating the long-term static risk profile of the healthcare portfolio, from the standard model specifications, shows consistent risk characteristics. The healthcare portfolio is relatively insensitive to the market cycle, with 𝛽 < 0.80. This is in good agreement with the prior literature, showing that the inelastic demand for healthcare products creates a lesser dependence on the market cycle. Furthermore, the healthcare portfolio has a modest but significant negative exposure to the size factor, indicating that the portfolio has a larger weight in less-risky larger companies. This is in line with the described change within the healthcare industry over the last decades, from many relatively small healthcare start-ups to a larger more mature industry. Next, the negative exposure to the HML factor shows that healthcare firms have a relatively high market valuation of equity. This reflects the believe by investors that healthcare companies will be able to benefit from profitable future opportunities,
19
lowering the required return. Last, in accordance to earlier findings by Tresl et al. (2014), Table 3 shows a small exposure to the momentum characteristic. To summarize, the overall healthcare portfolio can be characterized as consisting of slightly larger firms, with good expected growth potential and with small exposure to the momentum anomaly. Except for the small 𝛽 , the healthcare portfolio acts as insurance against the conventional risk factors, decreasing the return normally required by investors. However, this is not reflected in the ex-post realized returns, resulting in the large alphas.
As stated earlier, the inclusion of the policy uncertainty factor and the general uncertainty (control-)factor does not eliminate the ex-post alphas of the healthcare portfolio. However, the inclusion of both risk factors has two significant implications for the risk assessment of this portfolio.
First, the healthcare industry has an, at most marginally significant, positive relation to shocks in policy uncertainty. In other words, historical healthcare returns have shown to increase in conjunction with unexpected hikes in policy uncertainty, making the portfolio even less risky and lowering the required return. This finding is conflicting with the point-of-view by Koijen et al. (2016),
Table 5: Healthcare alphas from time-series regressions with uncertainty
The table reports time-series OLS estimates of 1) standard asset pricing models and 2) asset pricing models augmented with the factor mimicking portfolios of shocks in EPU and VXO. The models are specified as:
𝑟, = 𝑎 + 𝛽 𝑓 + ϵ
𝑟, is the monthly excess return on the healthcare industry portfolio, covering the period from 1986:02 to
2017:08. The risk factors 𝑓 are the excess return on the CRSP value-weighted market portfolio (Mkt) in case of the CAPM and the market excess return (Mkt), return on the zero-cost “small minus big” factor (SMB), the return on the zero-cost “high minus low” book-to-market factor (HML) in case of the FF3. The Carhart 4-factor model adds the return on the zero-cost, prior 2-12 month “up-minus-down” (UMD), momentum factor. The risk factors capturing economic policy uncertainty (EPU) and general uncertainty (VXO) are the zero-cost factor mimicking portfolios of the innovations in the Baker et al. (2016) EPU index and the month-end S&P 100 volatility. Alphas are reported as annualized percentages. The t-statistics are reported in parentheses and calculated using Newey-West (1987) adjusted HAC robust covariance matrices, with automatic lag length selection following Andrews (1991).
Standard Including uncertainty CAPM FF3 CAR4 CAPM FF3 CAR4 𝛼 (annual, %) 4.22** 4.85*** 3.97** 4.31** 4.62** 3.81** (2.17) (2.73) (2.26) (2.30) (2.59) (2.15) 𝛽 0.77*** 0.78*** 0.80*** 0.71*** 0.42* 0.47* (13.94) (15.40) (15.66) (2.90) (1.85) (1.94) 𝛽 -0.25*** -0.26*** -0.29*** -0.29*** (-4.36) (-4.82) (-4.41) (-4.66) 𝛽 -0.21* -0.17 -0.28*** -0.24** (-1.88) (-1.47) (-2.94) (-2.18) 𝛽 0.10 0.09 (1.45) (1.43) 𝛽 0.10* 0.07 0.08 (1.71) (1.25) (1.53) 𝛽 -0.17 -0.45* -0.43* (-0.65) (-1.90) (-1.74)
No. monthly obs. 379 379 379 379 379 379 adj. R2 0.54 0.58 0.59 0.55 0.59 0.60
20
who argue that government intervention can negatively affect cash flows and that this constant danger should require investors to command a premium. However, from the asset pricing literature, discussed briefly in section 3.2, it can also be argued that the expected part of government intervention risk is most likely already incorporated in the price of a security. Only the sensitivity to unexpected changes in this possibility will, theoretically, affect the required return by investors. The small positive relation between unexpected increases in policy-uncertainty and historical excess returns can potentially be explained by positive cashflow effects. Brogaard and Dretzel (2016) shows that the net dividend growth for the U.S. market as a whole, is unaffected by unexpected changes in policy uncertainty. However, a more detailed analysis might show industry-level differences. The market power of some healthcare firms combined with an inelastic demand, might enable them to take pre-emptive actions as a response to policy uncertainty shocks. Hence, further research could employ an event study approach, focussing on firm-level decision making and subsequent cashflow effects as a reaction to unexpected changes in policy uncertainty.
The second implication of including uncertainty factors is that the market risk factor from the standard asset pricing models is decomposed in a sensitivity with respect to the market cycle and a sensitivity to shocks in general economic uncertainty. The Fama-French and Carhart models shows a marginally significant and negative sensitivity of approximately -0.44 with respect to economic uncertainty. In both models, the market sensitivity decreases after inclusion of uncertainty variables from 0.80 to 0.47. The lower premium paid to compensate for market risk is most likely offset by the premium paid for unexpected changes in general economic uncertainty. Market cycle risk is estimated as the beta (𝛽 ) with respect to excess return on the CRSP value weighted index and the risk related to general economic uncertainty is given by the beta (𝛽 ) relative to a zero-cost mimicking portfolio of innovations in the S&P100 volatility index. Brogaard and Dretzel (2016) show that both factors measure different, but related concepts. Hence, to a certain extent this will increase concerns about multicollinearity within the model. However, this does not affect any inference from the alpha estimates. In addition, Brogaard and Dretzel (2016) also have shown that inclusion of VXO as a control variable does improve the model by effectively isolating the policy uncertainty from general uncertainty in the EPU factor.
4.1.2 – 36-month rolling-window estimates
The full-window regression estimates reject the hypothesis that the healthcare portfolio has a significant negative exposure to the EPU factor. Which is equivalent to stating that investors do not require an additional premium to compensate for policy uncertainty. However, prior literature has shown that fundamental characteristics of the healthcare industry have changed over time, directly affecting different sources of systematic risk. The time-evolution of the “augmented Carhart four-factor” regression estimates, as shown in figure 3, can therefore provide additional important insights. The figure plots trailing 36-month OLS estimates for the Carhart model with the uncertainty factors. The 36 month interval and regression methodology is adopted from Tresl et al. (2014), to allow for direct comparison.
All trailing 36-month estimates plotted in figure 3 are mean reverting, with all mean values significant at the 1% level. Moreover, the mean trailing 36-month alpha is found to be 33 basis points per month (3.97% per annum). This result further substantiates the main finding that inclusion of uncertainty factors does not explain the large excess-returns. This alpha estimate is similar in size to the full-window alpha of 3.81% per annum. The plot of time-variation in alpha shows that alphas measured exclusively in the window 2000-2008 will be close to zero, resulting in spurious inference about the true accuracy of the model in the long-run. In addition to this period of below-average alpha, also periods of persistent above-average alpha can be identified. For example, regressions over the 36
21
Figure 3: Trailing 36-month healthcare alphas and risk exposures
The figure shows time-varying OLS estimates of 36-month rolling-window regression of the Carhart 4-factor model (CAR4), augmented with the factor mimicking portfolios of shocks in EPU and VXO. The model is specified as:
𝑟, = 𝑎 + 𝛽 𝑓 + ϵ
𝑟, is the monthly excess return on the healthcare industry portfolio, covering the period from 1986:02 to
2017:08. The 6x1 matrix 𝑓 contains the risk factors Mkt, SMB, HML, UMD, EPU and VXO, as described in Table 5. Trailing 36-month estimates start from 1989:01 and are re-estimated monthly.
22
months before 1996, 2001 and 2014 yield positive alphas above 6% per annum and as high as 12%. Lastly, alpha estimates appear to have become more volatile since 1995.
The sensitivity with respect to innovations in the EPU index moves around zero and shows no clear trend over the sample period. The average trailing 36-month beta is 0.04 with a corresponding standard deviation of 0.14. From Figure 3 it can be seen that there are three distinct periods in which 𝛽 was below zero. More specific, the periods 1993-1996, 2006-2008 and 2012-2015 show a risk profile for the healthcare portfolio that is in line with the prediction by Koijen et al. (2016), that there should be a negative relation between excess returns and policy-uncertainty, causing investors to command a premium for holding these securities. However, the alphas estimated over the same periods don’t tend to zero. This shows that investors are still being significantly over- or undercompensated, even on the intervals that do satisfy the Koijen et al. (2016) model.
Figure 3 conveys three additional insights. First, the healthcare industry started to move in opposition to the market around 2000. This is shown in the time-evolution of exposure to the general market return (Mkt) and general uncertainty (VXO). This led prior research to conclude that the correlations have both decreased on average and become more volatile, suggesting that a larger allocation to healthcare will enhance portfolio diversification. However, it can be seen that the change was only temporal and that the market beta has reverted back to its original mean in the recent year. Second, the downward trend in size-related factor loading discussed in, among others, Golec and Vernon (2007) and Harrington and Miller (2009), seems to have reversed after the financial crisis in 2007-2008. Third, the drawdowns in the healthcare sector, as defined by Koijen et al. (2016), related to the Clinton healthcare reform in 1992-1993, the technology crash during 2000-2002 and the financial crisis in 2007-2008 appear to be followed by increases in the book-to-market factor loading. 4.1.3 – Summarizing main result
An answer to the research question whether compensation for increased exposure to policy-related economic uncertainty decreases the persistent excess returns on U.S healthcare equity, is twofold. First, the healthcare industry shows an insignificant or even a favourable sensitivity to shocks in policy-related economic uncertainty. As a result, any risk premium for this type of systematic risk will not explain the persistence of alpha in this industry. Second, even in sub-periods that do show signs of significant adverse sensitivity to shocks in policy uncertainty, the additional risk premium paid does not eliminate the estimated excess returns.
4.2 Healthcare sub-industry comparison of alphas and risk exposures
This section examines the robustness of the main result by evaluating differences on a sub-industry level. Prior literature shows that there are fundamental differences between the R&D dependent pharmaceutical and medical devices industries, on the one hand, and the health services industry on the other.
Table 6 shows static estimates of the standard and augmented asset pricing models. Panel A shows that the positive alphas on the healthcare portfolio originate predominantly from the pharmaceutical and medical devices sub-industries. The pharmaceutical portfolio earns the highest alpha of approximately 5% per annum and the medical devices portfolio earns a significant 4% risk-adjusted excess return. The health services industry earns an alpha in the range 0.15% to -2.38%, which is insignificantly different from zero. Panel B shows that the inclusion of the uncertainty factors does not change the view that large positive alphas are associated with the R&D intense industries. Furthermore, inclusion of the uncertainty risk factors results in an insignificantly small decrease in estimated alphas.
