Three Essays in Finance

Dissertation

Presented in Partial Fulfillment of the Requirements for the Degree Doctor of Philosophy in the Graduate School of The Ohio State University

By

Sehoon Kim, M.Sc.

Graduate Program in Business Administration

The Ohio State University

2017

Dissertation Committee:

Ren´eM. Stulz, Advisor Kewei Hou Bernadette A. Minton Berk A. Sensoy c Copyright by

Sehoon Kim

2017 Abstract

This dissertation is comprised of three essays that study the consequences of fi- nancial frictions in the context of corporate finance and asset pricing. The first and second chapters are empirical investigations of how firms use their internal finances to make real corporate decisions, and how external forces influence the efficacy with which firms utilize their internal resources. The third chapter is an attempt to quan- tify market frictions and their impact on the cross-section of stock returns.

In the first chapter, I study how corporate cash holdings impact firms’ product pricing strategies. Exploiting the Aviation Investment and Reform Act of the 21st

Century as a quasi-natural experiment to identify exogenous shocks to competition in the airline industry, I find that firms with more cash than their rivals respond to intensified competition by pricing more aggressively, especially when there is less concern of rival retaliation. Financially flexible firms based on alternative measures respond similarly. Moreover, cash-rich firms experience greater market share gains and long-term profitability growth. The results highlight the importance of strategic interdependencies across firms in the effective use of flexibility provided by cash.

The second chapter studies the effects of hedge fund activism on the activity and efficiency of target companies’ internal capital markets. I find that firms targeted by activist hedge funds significantly increase investment cross-subsidies between divi- sions, predominantly by enhancing the efficiency of their internal resource allocations.

ii Following Schedule 13D filings by activist hedge funds, segment investments of tar-

geted companies become more sensitive to cash flow generated elsewhere in the firm,

and this increase in cross-subsidization is primarily driven by the redirection of firm

cash flows toward segments with high Tobin’s Q. The increases in the activity and efficiency of internal capital markets due to hedge fund activism are unlikely to be driven by measurement errors in Tobin’s Q or changes in unobserved correlations across segments.

In the third chapter, co-authored with Kewei Hou and Ingrid Werner, we propose a parsimonious measure based solely on daily stock returns to characterize the sever- ity of microstructure frictions at the individual stock level and assess the impact of frictions on the cross section of stock returns. Based on our measure, stocks with the largest frictions command a value-weighted return premium as large as 10% per year on a risk-adjusted basis. The friction premium is stronger among small, low price, volatile, value, and illiquid stocks. Return spreads associated with momentum and idiosyncratic volatility are smaller and statistically less significant than previously documented after screening out stocks with high microstructure frictions. Using UK data, we show that our measure is useful in settings where the availability of quality data on trading volume, bid-ask prices, and intraday high-low prices is limited.

iii Dedicated to all my beloved ones.

iv Acknowledgments

This dissertation could not have come to fruition without the boundless support of those who deserve much more than these few words of gratitude.

First and foremost, I am deeply indebted to my advisor, Ren´eM. Stulz, who has guided, pushed, and supported me throughout the uncharted 5 years of my early academic life. I am proud to have joined a lineage of pupils who will always look up in awe and inspiration. I would also like to thank my dissertation committee, Kewei

Hou, Bernadette A. Minton, and Berk A. Sensoy, for their advice and encouragement.

I am particularly grateful to my co-authors of the third chapter, Kewei Hou and In- grid Werner, from whom I learned a great deal. I am also thankful to my colleagues for their comments and friendship. I also send a special thanks to Robyn Scholl, who has always gone beyond her way to ensure my academic and personal well-being.

My parents have greatly shaped my values, tenacity, and work ethic, which have made the eventual completion of my PhD possible. I would like to thank them for making me who I am, and for providing their unconditional love and support (not to mention patience). Also thanks to my sister for bearing with an increasingly pedantic older brother.

I thank my friends back home, Eung Kyu and Wonbin, for their unwavering friend- ship and loyalty through the best and toughest times. Finally, thank you Bohyun for bringing happiness and hope to this challenging adventure.

v Vita

1984 ...... Born in Seoul, South Korea

2008 ...... Summer Analyst, Goldman Sachs

2009 ...... B.B.A. in Business Administration, Seoul National University

2012 ...... M.Sc. in Economics, London School of Economics

2012 to present ...... Graduate Research Associate, Department of Finance, The Ohio State University

Fields of Study

Major Field: Business Administration

Area of Specialization: Finance

vi Table of Contents

Page

Abstract...... ii

Dedication...... iv

Acknowledgments...... v

Vita...... vi

List of Tables...... ix

List of Figures...... xi

Chapter 1. Cash, Financial Flexibility, and Product Prices - Evidence from a Natural Experiment in the Airline Industry...... 1

1.1 Introduction...... 1 1.2 Identification Strategy...... 11 1.2.1 The Endogeneity Problem...... 11 1.2.2 Aviation Investment and Reform Act for the 21st Century. 13 1.3 Data and Empirical Specification...... 20 1.3.1 Data Sources and Screening...... 20 1.3.2 Variables...... 24 1.3.3 Empirical Specification...... 29 1.4 Results...... 31 1.4.1 Main Results...... 31 1.4.2 Subsample Analyses...... 36 1.4.3 Do Firms Hold Cash Anticipating AIR-21 Coverage?.... 41 1.4.4 Is the AIR-21 Discontinuity Spurious?...... 43 1.4.5 Are Cash Rich Firms Constrained?...... 45 1.4.6 Cash and Market Performance Outcomes...... 50

vii 1.4.7 The Effects of Competition on Corporate Cash Holdings.. 53 1.5 Conclusion...... 55

Chapter 2. Hedge Fund Activism and Internal Capital Markets...... 59

2.1 Introduction...... 59 2.2 Empirical Strategy...... 65 2.2.1 Data...... 65 2.2.2 Endogeneity...... 68 2.2.3 Methodology...... 71 2.3 Key Results...... 75 2.4 Conclusion...... 87

Chapter 3. (Priced) Frictions...... 90

3.1 Introduction...... 90 3.2 Simple Model...... 96 3.2.1 Basic Setup...... 96 3.2.2 Methodology: Constructing “FRIC”...... 98 3.3 Data...... 101 3.4 Characteristics of Friction-Sorted Portfolios...... 104 3.5 Cross-Sectional Return Predictability: Portfolio Sorts...... 111 3.5.1 Returns of FRIC Portfolios...... 111 3.5.2 Robustness...... 113 3.6 Cross-Sectional Return Predictability: Fama-MacBeth Regressions 116 3.7 Interactions between FRIC and Firm Characteristics...... 122 3.8 Anomalies and Frictions...... 125 3.9 Extension: Results for the United Kingdom...... 133 3.10 Conclusion...... 139

Bibliography...... 142

Appendix A. Additional Figures and Tables to Chapter 1...... 152

Appendix B. Data and Variable Descriptions to Chapter 2...... 158

B.1 Variables...... 158 B.2 Data Screening...... 159

Appendix C. Addendum to Chapter 3...... 162

C.1 Adjusting FRIC for Serial Correlation in True Returns...... 162

viii List of Tables

Table Page

1.1 Airline and Market Summary Statistics...... 23

1.2 Airlines and Airports...... 26

1.3 Key Variable Statistics...... 28

1.4 Cash, Competition, and Multimarket Contact...... 33

1.5 Low Cost Carriers vs. Legacy Airlines...... 38

1.6 Ex-Ante Market Dominators vs. Laggards...... 40

1.7 Cash 3 to 4 Years Prior, Competition, and Multimarket Contact... 42

1.8 Placebo Tests with Alternative Treatment Thresholds...... 44

1.9 Financial Flexibility vs. Constraints...... 47

1.10 Cash and Market Performance Outcomes...... 51

2.1 Sample Statistics...... 67

2.2 Hedge Fund Activism and Internal Capital Markets...... 76

2.3 Internal Capital Market Efficiency: Alternative High Q Definitions.. 80

2.4 Matched Sample Difference-in-Differences Regressions...... 82

2.5 Hedge Fund Activism and Firm diversification...... 85

ix 3.1 Characteristics of Friction-Sorted Portfolios...... 105

3.2 Microstructure Frictions and the Cross-Section of Expected Stock Re- turns: Portfolio Sorts...... 112

3.3 Microstructure Frictions and the Cross-Section of Expected Stock Re- turns: Fama-Macbeth Regressions...... 117

3.4 Friction Premiums across Characteristic Quintiles...... 123

3.5 Momentum, IVOL, and FRIC...... 127

3.6 Illiquidity and the Cross-Section of Stock Returns in the United Kingdom135

A.1 Passenger-Weighted Relative-to-Rival Cash and Multimarket Contact 154

A.2 Key Variable Statistics for Select Subsamples...... 155

A.3 Alternative Fixed Effects and Standard Error Clustering...... 156

A.4 Difference of Means: Control vs. Treated Groups...... 157

C.1 Microstructure Frictions and the Cross-Section of Expected Stock Re- turns: Addressing Serial Correlation in True Returns...... 166

C.2 Fama-Macbeth Regressions (WLS)...... 167

x List of Figures

Figure Page

1.1 Industry Financial Constraints...... 3

1.2 The Impact of Cash on Pricing across Multimarket Contact Terciles.7

1.3 Research Design...... 12

1.4 Discontinuity around the AIR-21 Treatment Threshold...... 15

1.5 No Anticipatory Adjustment of Cash Holdings...... 17

1.6 No Discontinuity in Airport-Level Control Variables...... 19

1.7 Cash Holdings Before and After AIR-21 Coverage...... 54

3.1 Time-Series of Market Average Friction Estimates...... 109

A.1 Correlations Between Explanatory and Control Variables...... 153

xi Chapter 1: Cash, Financial Flexibility, and Product Prices - Evidence from a Natural Experiment in the Airline Industry

1.1 Introduction

At the end of 2015, U.S. non-financial companies held some $2 trillion in cash, nearly twice the amount half a decade ago, heightening interest among finance aca- demics and policy makers alike.1 In light of the era of large corporate cash holdings, recent studies have argued that cash can be a valuable source of financial flexibility for firms (see Gamba and Triantis(2008), Denis and Sibilkov(2010), Denis and McK- eon(2012)). An important avenue through which this flexibility can prove valuable is product market dynamics. A number of papers have recently shown that cash holdings have a large positive impact on market share outcomes (see Fresard(2010)), and that product market threats significantly influence cash retention decisions (see

Hoberg, Phillips, and Prabhala(2014)). This study goes deeper to explore whether and when companies tap into their cash war chests to formulate competition strate- gies to their advantage. Do firms with larger cash holdings relative to their peers price their products more aggressively, consistent with ‘Long purse’ arguments (see Telser

1J.P. Morgan Chase & Co. reported that non-financial companies in the S&P 500 had nearly $2.1 trillion in cash at the end of October, 2015. At the end of 2014, Standard & Poor’s Ratings Services had reported that around 2,000 rated U.S. non-financial companies held $1.82 trillion in cash, and Moody’s had reported a similar amount of $1.73 trillion.

1 (1966), Bolton and Scharfstein(1990))? When in particular? In a novel empirical investigation of the airline industry incorporating the interdependence of competition strategies across firms, this paper shows that cash provides financial flexibility which enables firms to undercut their competitors, predominantly when they face less po- tential retaliation from rivals.

To capture potential rival retaliation, I draw from the industrial organization lit- erature the idea of multimarket contact and mutual forbearance. Firms very often serve several markets, for example by having multiple product lines or operating across geographical segments. In such a multimarket setting, altering strategy in one market can affect the actions of rival firms in other markets due to the inter- connection of competition strategies across markets (see Bulow, Geanakoplos, and

Klemperer(1985)). Bernheim and Whinston(1990) expand on this idea and argue that competitors who encounter more frequently due to broader market overlap (i.e. higher multimarket contact) recognize the interdependence of their strategies, and are more likely to collude in equilibrium (i.e. engage in mutual forbearance) for fear of what rival firms might do in other jointly contested markets. Evans and Kessides

(1994) show that airlines indeed live by the ‘golden rule’ where they refrain from initiating aggressive pricing actions when multimarket contact is high. Hence, I take the multimarket contact measure motivated by IO theory to capture potential rival retaliation concerns, and form the hypothesis that higher multimarket contact should dampen the strategic benefit of cash predicted by financial economic theory. In this paper, I empirically investigate whether cash-rich firms price more aggressively and whether multimarket contact weakens this competitive role of cash.

There are two main challenges to this analysis. First is that there need be a set-

2 KZ Indices WW Indices SA Indices

0 10 20 30 40 50 0 10 20 30 40 50 0 20 40 60 80

Figure 1.1: Industry Financial Constraints

This figure compares the average KZ index (following Kaplan and Zingales(1997), Lamont, Polk, and Sa´a-Requejo(2001)), WW index (following Whited and Wu(2006)), and SA index (following Hadlock and Pierce(2010)) of firms in the airline industry (SIC code 4512) with firms in Fama and French(1997) 48 industries. Using the Compustat universe of firms, the KZ index is computed for each firm as -1.002×Cash flow+0.283×Tobin’s Q+3.139×Debt−39.368×Dividends−1.315×Cash, where cash flow is oibdp/at, Tobin’s Q is market value of assets (at + csho × prcc f − ceq − txdb) divided by 0.9×book value of assets (at)+0.1×market value of assets, debt is (dlc + dltt)/at, dividends are (dvc + dvp)/at, and cash is che/at. The WW index is computed for each firm as -0.091×Cash flow−0.062×DIVPOS+0.021×Long-term debt−0.044×Log assets+0.102×Industry sales growth−0.035×Sales growth, where DIVPOS is an indicator variable for whether the firm pays dividends and long-term debt is dltt/at. The SA index is computed for each firm as - 0.737×Size+0.043×Size2−0.040×Age, where size is the log of Min(at, $4.5 billion) and age is Min(Firm age, 37 years). Each year, firms are ranked into 1/100th percentiles based on their KZ, WW, and SA indices. Then, the KZ, WW, and SA ranks are averaged across firms in the same industry. Finally, the time-series averages of the industry KZ, WW, and SA ranks are presented on a scale of 1 to 100.

ting where rivalry and markets are cleanly defined, and second is that both cash and market overlap are likely to be endogenously linked to firm pricing behavior, making it difficult to make causal inferences. I overcome both of these issues by focusing

3 on the airline industry. Taking directional air routes as markets, defining rivalry is a simple and clean task in this industry since route services are comparable across airline companies, and rich data on ticket prices serve as a readily available source of market pricing information.2 The airline industry is also an appropriate place to study the impact of cash holdings in the sense that it is an industry with relatively high financial constraints where financial distress and bankruptcies associated with borrowing constraints have frequented headlines throughout recent decades.3 Fig- ure 1.1 charts the average KZ index (following Kaplan and Zingales(1997), Lamont,

Polk, and Sa´a-Requejo(2001)), WW index (following Whited and Wu(2006)), and

SA index (following Hadlock and Pierce(2010)) of firms in the airline industry in comparison with Fama and French(1997) 48 industries. The indices show that air- line companies on average are more constrained than firms in many other industries, indicating that cash holdings should play a role in their corporate decisions.4 The relevance of corporate cash holdings for pricing competition in the airline industry is very real. At the end of 2015, stated its intent to engage in a price war with low cost carrier rivals such as , which coincided with its CFO’s announcement that the company had “more cash than we need at this time” and was widely purported to be due to ample financial slack in times of low fuel prices.

The empirical design of this study allows me to effectively sidestep the endogeneity

2This has led a number of recent studies in finance to rely on the airline industry. For example, Azar, Schmalz, and Tecu(2016) focus on the airline industry to study the effects of common own- ership on competition, while Parise(2017) uses the industry to demonstrate that potential changes in competition dynamics can influence the debt structure decisions of firms ex-ante. 3See Weiss and Wruck(1998) for a detailed study of the famed bankruptcy case of Eastern Airlines. 4In a frictionless Modigliani and Miller(1958) world where firms can freely borrow from the external capital market, cash holdings should have no bearing on the firm’s policies.

4 problem. Cash is measured for each firm in comparison with its rivals’ in each market, such that the amount of relative-to-rival cash is not self-selected solely by the firm but determined in conjunction with the choice of its competitors, as is multimarket contact by construction. On top of that, I exploit an industry wide regulation, the

Aviation Investment and Reform Act for the 21st Century (AIR-21), as a quasi-natural experiment to identify plausibly exogenous shocks to market-level competition and infer how ex-ante relative-to-rival cash holdings and multimarket contact prior to such shocks affect ex-post pricing. Under AIR-21, airports of certain size whose two largest airlines board more than 50% of the airport’s total passengers are required to submit competition enhancement plans to the Federal Aviation Administration (FAA) and implement them under periodic FAA monitoring (detailed in the following section).

Therefore, AIR-21 serves as a competition shock to a market with a covered airport at either endpoint (e.g. the origin). The design of this regulation ensures exogene- ity of these shocks in two ways. First, assessment of the treatment effect can be made just around the 50% top-two airline concentration ratio threshold, facilitating a regression discontinuity (RD) approach which circumvents concerns regarding large unobservable differences between treated and non-treated markets. Second, AIR-21 coverage for a given year is determined by passenger enplanement of the two most dominant airlines relative to airport totals based on data from two years prior. It is thus unlikely that an airline would be able to manipulate boardings in a way that purposefully affects AIR-21 coverage of an airport.

With this setting, I implement a triple difference framework with a flavor of regres- sion discontinuity (RD) design on firm-market and quarter panel data. Specifically, I run regressions of changes in pricing strategy on ex-ante relative-to-rival cash, ex-ante

5 multimarket contact, and AIR-21 treatment, including a host of firm and market-level control variables as well as firm, market, and time fixed effects. Notably, I explicitly control for relative-to-rival debt and its interactions with AIR-21 and multimarket contact to tease out the impact of cash distinct from the effect of leverage well known in the literature. To apply a quasi-RD framework, I run regressions on progressively narrower windows around the 50% treatment cutoff of endpoint airports and show that results are robust, if not stronger, in closer regions surrounding the threshold.

The main results confirm that firms with larger cash holdings relative to their rivals respond to market-level competition shocks by pricing more aggressively, but only when multimarket contact is sufficiently low (i.e. when there is less concern of retaliation from rivals). In economic magnitudes, a one standard deviation or approx- imately 9 percentage point increase in relative-to-rival cash as a fraction of assets one year prior to an AIR-21 competition shock in a market leads to roughly 15 percent- age points lower price growth over the next 36 months compared to the previous 36 months (i.e. half standard deviation lower price growth differential). Multimarket contact has a sizable impact on this strategic effect of cash: a 5% increase would almost overturn the cash effect.

Figure 1.2 provides a snapshot of the main results, where pricing to cash holding sensitivities are charted across multimarket contact terciles, based on a sample close to the AIR-21 treatment threshold. The sensitivities are obtained from coefficients of the interaction terms in difference-in-differences regressions of changes in pricing strategy on AIR-21 treatment and relative-to-rival cash. It can be seen that the competitive effect of cash, namely that greater cash reserves enable firms to price

6 0.3

0.2

0.1

0

-0.1

-0.2

-0.3

-0.4

-0.5

-0.6 MC tercile 1 MC tercile 2 MC tercile 3

Price Growth-Cash Sensitivity Unconditional Sensitivity

Figure 1.2: The Impact of Cash on Pricing across Multimarket Contact Terciles

This figure illustrates the main results of the paper. Price growth differential to cash holding sensitivities are charted across multimarket contact terciles. Each period, firm-market observations are sorted into terciles based on multimarket contact. Within each multimarket contact (MC) tercile group, price growth differential to cash holding sensitivity is obtained as the coefficient on the interaction term between AIR-21 treatment and relative-to-rival cash holding in a difference-in- differences regression of change in pricing strategy (price growth over the next 36 months compared to the previous 36 months) on AIR-21 treatment and relative-to-rival cash to assets ratio. The unconditional sensitivity across all multimarket contact (MC) terciles is shown as the dotted line. Results are based on a restricted sample where 2-year prior top-two airline concentration ratios at market origin airports are 10% above and below the 50% AIR-21 treatment cutoff. Variable constructions are detailed later in Section 1.3.

aggressively, is pronounced when there is less concern of rival retaliation and is at- tenuated, even reversed, as multimarket contact increases.

In addition, I exploit heterogeneity in fare levels and market shares across firms to show that the effects of AIR-21 on different firms are consistent with what would be

7 expected given the nature of the regulation. Arguably, AIR-21 treatment should have differential effects across firms in the same market since the aim of its legislation is to lower prices and distribute passenger boardings more evenly across airlines. For in- stance, Snider and Williams(2015) show that AIR-21 led to lower airline fares mainly through gate reallocations toward entrant low cost carriers (LCCs). Consistent with these implications, I find that LCCs respond to AIR-21 competition shocks by pricing aggressively irrespective of their cash holdings, while legacy airlines respond aggres- sively conditional on holding more cash. In both cases, their responses are dampened by higher multimarket contact. Also, the main results hold only for firms that had high ex-ante market share (i.e. firms for which AIR-21 indeed serves as a competition shock), but are non-existent for firms that had low market share to begin with (i.e.

firms for which AIR-21 rather serves as an accommodative event).

I further provide evidence from a number of robustness checks. To alleviate con- cerns that firms might predict AIR-21 coverage and build-up cash reserves in advance,

I use relative-to-rival cash measured 3 and 4 years prior to treatment (one year prior in baseline specifications) and show that results are robust. Placebo tests using al- ternative threshold levels of top-two airline passenger shares as AIR-21 treatment cutoffs, 40% and 60% instead of the baseline 50%, confirm that the main results are unlikely due to other confounding effects that happen to coincide with competition shocks induced by AIR-21.

To cement the argument that the effect of cash holdings on price growth differ- entials is that of financial flexibility, I show that high net cash or high payout firms compete aggressively in response to AIR-21 as do cash-rich firms, in contrast to the

8 opposite accommodating behavior of supposedly constrained firms that had cut divi- dends in the previous year. I also demonstrate that the market performance outcomes of holding more cash than rivals, i.e. market share gains and long-term profitability growth, are consistent with cash being a valuable source of financial flexibility. Fi- nally, I show suggestive evidence that AIR-21 competition shocks lead to increased corporate cash holdings, which is consistent with firms rationally responding to in- tensified competition by building up financial war chests to use for aggressive pricing.

This paper contributes to the growing literature studying the interaction of fi- nancial flexibility and product market competition. The predominant approach to understanding the relationship between financial strength and competition is based on ‘long purse’ or ‘deep pocket’ arguments (see Telser(1966), Bolton and Scharfstein

(1990)). Under this approach, a weak balance sheet (e.g. little cash, high leverage) takes away the ‘long purse’ from firms, rendering them unable to price aggressively and prone to forgo future market shares for profits today. This line of argument has gained empirical support by papers relating financing decisions to competition, notably by Chevalier(1995a, 1995b) in the context of leveraged buyouts in the su- permarket industry, and Campello(2003) who provides more general evidence that levered firms raise prices during recessions to maximize short-term profits when rivals are also levered, consistent with markup counter-cyclicality theories `ala Chevalier and Scharfstein(1996). In a similar spirit but focusing on capacity investments in the casino industry rather than pricing, Cookson(2017) also shows that high leverage prevents firms from responding to competition threats. Parise(2017) conversely shows that firms in the airline industry increase their debt maturities in the face of entry threats to lower roll-over risk. Only more recently has this ‘deep pocket’ story been

9 tied to the precautionary motive for cash. Haushalter, Klasa, and Maxwell(2007)

find that the extent to which firms have interdependent growth prospects with rivals

(i.e. face higher predation risk) is positively associated with cash holdings and the use of hedging derivatives. Fresard(2010) uses shifts in import tariffs as exogenous competition shocks and shows that firms with large cash reserves beforehand gain larger market shares ex-post. This paper adds more color to recent developments in the literature by documenting how cash affects firm pricing strategies, thereby shed- ding light on the mechanism through which cash impacts product market outcomes.

At its roots, this paper is part of a vast literature on corporate cash policy. The precautionary saving motive for cash argued by Keynes(1936) has been studied by numerous papers. Fazzari and Petersen(1993) find that firms use working capital

(e.g. cash) to smooth fixed investments. Opler, Pinkowitz, Stulz, and Williamson

(1999) and Bates, Kahle, and Stulz(2009) suggest cash flow volatility to be a key de- terminant of corporate cash holdings. Consistent with the insight of Modigliani and

Miller(1958) that financial slack should matter only when there are financing frictions,

Almeida, Campello, and Weisbach(2004) show that only financially constrained firms accumulate cash out of their cash flows. Faulkender and Wang(2006) also show that the marginal value of cash is greater when firms are financially constrained. Acharya,

Almeida, and Campello(2007) point out the hedging role of cash when cash flows are low and investment opportunities are high. In the recent 2007-2008 financial cri- sis, Duchin, Ozbas, and Sensoy(2010) show that cash serves as a buffer to supply shocks to external financing. How these cash reserves are used in firms’ day-to-day operations are of ever growing interest inside and outside of academia, and this paper furthers our understanding of such matters.

10 The remainder of the paper is organized as follows. In Section 1.2, I discuss the identification strategy in greater detail. Data, variables, and the empirical specifica- tion are described in Section 1.3. The main results of the paper and robustness tests are presented in Section 1.4. Finally, I conclude in Section 1.5.

1.2 Identification Strategy

1.2.1 The Endogeneity Problem

In this section, I elaborate on the background of my research design and identifi- cation strategy. The ideal method to study the effect of cash holdings on firm pricing strategy would be to take a firm-market observation, duplicate it, treat only one of them with a shock to cash, and then examine how each of their pricing policies subse- quently evolve. To test the role of multimarket contact in weakening or strengthening the strategic effect of cash, one would extend the duplication and treatment exercise this time with a shock to multimarket contact. Figure 1.3 illustrates such a research design.

There are obvious challenges to this ideal approach. Not only am I incapable of observing exact counterfactuals for treated observations, but I am also unable to change firm-market characteristics at random. This means that my explanatory vari- ables, cash holdings and multimarket contact, would likely be endogenously linked to the outcome variable of interest, firm pricing strategy, via some unobserved factor that is also correlated with the explanatory variables, causality flowing in the other direction, or other selection biases at play. Establishing causality from cash and mul- timarket contact to pricing is therefore a critical challenge of this study.

The framework of my analysis minimizes this issue in a number of ways. To start

11 Figure 1.3: Research Design

This figure illustrates the idea underlying the research design of this paper. Ideally, a researcher would take a firm-market observation, duplicate it, treat one of them with a shock to cash holdings, then look at how pricing policies evolve for each firm-market. A similar extension would be done with a shock to multimarket contact to study its effect on the strategic use of cash. The methodology of this paper allows me to accomplish an analysis close to this ideal by treating firm-markets with AIR- 21 competition shocks while using relative-to-rival cash and multimarket contact prior to treatment as sources of ex-ante variation. The large number of firm-markets facilitates precise estimation of the treatment effect on prices. I control for unobserved firm-market invariants by including firm and market fixed effects, and also control for other observed variables to account for possible differences between treated and non-treated groups.

with, the relevant amount of cash in this study is what the firm has in excess of what its rivals have. This variable, which I refer to as relative-to-rival cash, is determined not by the firm alone but jointly by the firm and its rivals. It therefore varies across markets for a given firm, and suffers less from self-selection than would firm-level cash holdings. In a similar vein, multimarket contact is also a jointly determined variable over which the firm does not have complete control. That said, this does not ensure

12 that relative-to-rival cash and multimarket contact are exogenously given. Another

device is needed to fully address endogeneity.

1.2.2 Aviation Investment and Reform Act for the 21st Cen- tury

This endogeneity problem is circumvented by exploiting exogenous changes in the

competitive environment of markets, and studying the firm’s ex-post pricing policy

responses with respect to ex-ante variations in relative-to-rival cash and multimarket

contact. These exogenous changes are identified by use of an industry-wide regula-

tion that was legislated at the end of 2000, and went into effect at the beginning of

2001. The Wendell H. Ford Aviation Investment and Reform Act for the 21st Cen-

tury (AIR-21) called into question anti-competitive practices at airports, and required

those above a certain concentration level to undergo concrete procedures to make sure

entrant airlines could access airport facilities.5 Each year, large commercial airports

(i.e. those that enplane more than 0.25% of total U.S. passengers) are subject to coverage by AIR-21 if the two most dominant airlines control more than 50% of those airports’ passenger boardings, based on boarding data from two years prior. Covered airports are required to file detailed competition enhancement plans with the Federal

Aviation Administration (FAA) and the Department of Transportation (DOT), and are otherwise not approved of their Passenger Facility Charges (PFC) and Airport

Improvement Program (AIP) grants, which have been shown to comprise the bulk

5A Federal Aviation Administration (FAA)/Optimal Solutions & Technologies (OST) task force study “Airport business practices and their impact on airline competition” conducted in 1999 states that access at many of the nation’s most heavily used airports are limited due to business practices that prevent entry by new airlines or hinder competition among incumbent airlines, such as long- term exclusive-use gate lease agreements. The study’s recommendation of a competition enhancing policy was directly linked to the inclusion of competition plan requirement provisions in AIR-21.

13 of airport capital funding.6 Airports are subsequently required to submit two status

updates at 18 month intervals after initial coverage that show significant progress in

implementing their competition plans, giving them a total of 36 months until they

are potentially re-treated as a covered airport.7 Taking advantage of this regulation, markets that originate from airports treated by AIR-21 are then identified as those experiencing shocks to competition.

Using AIR-21 to identify shocks to market-level competition addresses the endo- geneity issue in two ways. To begin with, because AIR-21 coverage of an airport in a given year is determined by the passenger boarding share of the two largest air- lines based on data from two years prior, it is unlikely that airlines could manipulate boardings relative to their rivals’ such that they purposefully influence AIR-21 treat- ment to an airport two years down the road. More importantly, the regulation allows a regression discontinuity (RD) approach in which the local treatment effect of com- petition shocks can be measured just around the 50% top-two airline concentration ratio cutoff. This mitigates concerns about large unobservable differences between treated and non-treated firm-markets, so long as there is an apparent discontinuity in market-level competition around the threshold which arguably has little to do with conditions other than the AIR-21 treatment rule itself.

Figure 1.4 demonstrates the impact of AIR-21 on competition, specifically the dis- continuities in various measures of changes in competition around the 50% treatment

6A 2009 study “Airport capital development costs” conducted by Airports Council International North America (ACI-NA) finds that PFCs (21.7%), PFC backed bonds (30%), and AIP grants (22.2%) comprise the bulk of airport capital funding for committed projects. 7In a program guidance letter sent by the FAA to commercial airports, the following, among other details, are required to be addressed by filed competition plans: availability of gates and related facilities, leasing and subleasing arrangements, gate assignment policy, and gate use requirements. This indicates that gate reallocations are among the primary tools at the disposal of airports to enhance competition.

14

All Fares LCC Fares Legacy Fares LCC Share 50% Cutoff Threshold -0.073** -0.064** -0.048 0.059** (0.031) (0.030) (0.033) (0.027) Alternative Thresholds 40% Threshold -0.009 0.004 -0.024 -0.030 (0.039) (0.043) (0.043) (0.025)

60% Threshold -0.022 -0.021 -0.034 -0.009 (0.025) (0.030) (0.029) (0.021)

Observations 259 238 258 259

Figure 1.4: Discontinuity around the AIR-21 Treatment Threshold

These figures plot airport-quarter level observations of differences in average price growth between the next and previous 12 quarters for all airlines; low cost carriers (LCCs) only; legacy airlines only, and changes in LCC passenger share growth, with respect to the AIR-21 forcing variable, i.e. total passenger share of the top-2 airlines at each airport. Average prices and passenger shares are based on markets originating from each airport. The fitted lines and confidence bands are from local linear regressions with triangle kernels on each side of the 50% forcing variable treatment cutoff. The accompanied table shows results from non-parametric RD estimations via local linear regressions with triangle kernels and Imbens and Kalyanaraman(2012) optimal bandwidths. Placebo results are shown for alternative cutoffs (40% and 60%). Standard errors are shown in parentheses (*** p<0.01, ** p<0.05, * p<0.1).

15 threshold at market origin airports. In scatter plots of airport-quarter observations, I map average price growth differentials (top left), average low cost carrier (LCC) price growth differentials (top right), average legacy airline price growth differentials (bot- tom left), and LCC passenger share growth differentials (bottom right) with respect to the AIR-21 treatment forcing variable, the total passenger share of the top-two airlines at each airport. To take into account the accumulative 36 month period cov- ered airports are given to foster competition, growth differentials are computed as the difference of growth rates in the next 36 months and the past 36 months. The

fitted lines and confidence bands are from local linear regressions with triangle ker- nels on each side of the 50% forcing variable cutoff. The top left plot reveals a clear downward discontinuity in average price growth differentials, confirming that AIR-21 has an impact on price competition. This reduction in price growth is evident for

LCCs (top right) but not for legacy airlines (bottom left) and coincides with a dis- crete increase in LCC passenger share growth (bottom right), consistent with Snider and Williams(2015) where they find AIR-21 to have impacted competition mainly through gate reallocations toward entrant LCCs. The accompanying table shows the discontinuities more formally with results from non-parametric RD estimations via local linear regressions with triangle kernels and Imbens and Kalyanaraman(2012) optimal bandwidths. The local treatment effect of AIR-21 around the 50% threshold manifests in 7.3% lower average price growth, 6.4% lower LCC price growth, and

5.9% higher LCC passenger share growth. To show that these discontinuities are not simply the product of statistical coincidence due to confounding effects that happen to occur around the threshold, placebo results are shown for alternative cutoffs. At arbitrary threshold top-two airline concentration levels of 40% and 60%, there are no

16

Figure 1.5: No Anticipatory Adjustment of Cash Holdings

This figure plots airport-quarter level observations of ex-ante (lagged by 1 year) airport-level average airline cash holdings (scaled by total assets) with respect to the total passenger share of the top-2 airlines at each airport. All variables are based on markets originating from each airport. The fitted lines and confidence bands are from local linear regressions with triangle kernels on each side of the 50% forcing variable treatment cutoff.

such discontinuities, strengthening the interpretation that the estimated local treat- ment effects are due to AIR-21 rather than omitted factors that arise arbitrarily at different regions of the forcing variable.

A remaining concern is that firms may be able to predict AIR-21 treatment to markets and adjust their cash holdings in anticipation of such shocks, rendering cash holdings endogenous. This is unlikely to be a concern. First of all, it is highly implau- sible that firms could adjust their relative-to-rival cash holdings in each and every market they serve (i.e. at the firm-market level), especially given that the average

firm operates in roughly 750 markets. Their cash holdings are determined primarily at the firm level, while AIR-21 shocks occur at the market level. Moreover, there is no evidence that firms adjust firm-level cash holdings prior to AIR-21 treatment

17 to markets. Figure 1.5 plots the ex-ante average cash holdings of airlines at each airport measured one year earlier, against the top-two airline passenger share: the

AIR-21 forcing variable. Clearly, there is no discontinuity or bunching around the

50% threshold, indicating that changes in the forcing variable do not lead firms to preemptively reconfigure their cash holdings in the region near the cutoff. Although values of the forcing variable that are far away from the 50% cutoff, for example

30% or 70%, might have firms believe with little doubt that the likelihood of AIR-21 treatment is low or high, Figure 1.5 largely suggests that the treatment outcome is much less predictable just around the 50% threshold.

Figure 1.6 further solidifies the exogeneity of AIR-21 competition shocks by demonstrating that there are no ex-ante observable differences between treated and non-treated airports, especially around the 50% treatment threshold. Each subfigure plots various ex-ante airport-level control variables measured one year prior such as the total number of passenger boardings, average route distance, total number of originating routes, average market prices, average LCC prices, and average legacy airline prices against the top-two airline concentration ratio (i.e. the AIR-21 forcing variable). In contrast to ex-post changes in competition, there are no visible discon- tinuities in these variables around the 50% cutoff.

In short, AIR-21 provides a clean exogenous shock to competition in markets that originate from treated airports.8 Next, I describe the data, variables, and the triple difference empirical specification.

8In untabulated analysis, I find similar results for destination airport-level plots and RD estima- tions.

18 19

Figure 1.6: No Discontinuity in Airport-Level Control Variables

These figures plot airport-quarter level observations of ex-ante (lagged by 1 year) airport-level total number of passengers, average route distance, total number of routes, average market prices, average low cost carrier prices, and average legacy airline prices, with respect to the total passenger share of the top-2 airlines at each airport. All variables are based on markets originating from each airport. The fitted lines and confidence bands are from local linear regressions with triangle kernels on each side of the 50% forcing variable treatment cutoff. 1.3 Data and Empirical Specification

1.3.1 Data Sources and Screening

The data used in this study can be found at the Bureau of Transportation Statis- tics (BTS) and is comprised of three parts. The first part is firm-market pricing data, which is taken from the Airline Origin and Destination Survey (DB1B) database. The

DB1B database is a quarterly 10% random sample of all passenger-level itinerary pur- chases, domestic and directional. Each itinerary purchase observation has information on the airline’s identity (firm), origin and destination of the itinerary (market), per- passenger fare paid, number of passengers, any connections made, flight distance, whether it is part of a round-trip purchase, and whether it is an interline or online ticket. Markets are pre-defined in the DB1B database by the BTS as directional routes without ‘trip breaks’, which are points in an itinerary where the passenger is assumed to have stopped for reasons other than changing flights. For example, an itinerary BOS-LAS-BOS would have two markets BOS-LAS and LAS-BOS, with a trip break occurring at LAS. Firm-market level prices are computed by taking the quarterly passenger weighted-average of itinerary-level prices. These quarterly aver- age prices are calculated separately for indirect and direct flights, where round-trips are considered to be two equally priced one-way trips.

The second part is passenger boarding data used for identifying airport AIR-21 coverage, gathered from the T-100 Segment database. Unlike the DB1B database, the T-100 Segment database represents the full sample of all flights with at least one endpoint including a U.S. airport. It contains monthly data on airline identity, origin and destination of flight, and number of passengers. Because T-100 Segment data does not include information on on-demand and in-transit passengers, which are used

20 by the FAA to determine AIR-21 coverage, I supplement this with annual airport-

level all-enplanement data provided by the FAA.9

Finally, firm financial data are obtained from Schedule B-1 of Form 41 Financial

Data. Balance sheet information is available at quarterly frequency for large airlines

with annual operating revenues exceeding $20 million, and at semi-annual frequency

for smaller airlines. Virtually all airlines in the final sample report quarterly financial

data. Cash holdings and control debt variables are computed as a fraction of assets.

The sample period is from 2001Q1 to 2014Q4.

Several standard screening procedures are implemented on the data. To remove

effects of abnormally low or high prices due to ticket punching errors or mileage re-

demptions, I follow Snider and Williams(2015) and drop itineraries with prices lower

than $25 or higher than $2,500 in 2008 dollars. I also drop itineraries with more than

4 connections for round-trips, and those with more than 2 connections for one-way

trips. Itineraries associated with interline tickets or open-jaw travel are removed as

well.10 To determine whether an airline serves a particular market, I follow previous papers (see Berry(1992), Ciliberto and Tamer(2009)) and require that at least 90 passengers appear in the DB1B 10% random sample each quarter for a given firm- market. Also, routes are required to have at least 180 passengers in the DB1B sample each quarter to satisfy as a market. I focus the analysis on markets where both end- points are airports with passenger boardings above 0.25% of the U.S. total, which is the requirement for an airport to be considered for potential coverage by AIR-21.

9Though on-demand flights (e.g. charter flights) and in-transit passengers (e.g. passengers re- maining on planes stopping to refuel) comprise but a small number, they are important for facilitating a regression discontinuity analysis. 10Interline tickets refer to itineraries where different segments or legs of the travel are served by different airlines. Open-jaw travels are round-trips where the destination and/or origin are not the same in both directions.

21 Financial variables are winsorized at the bottom and top 1% to handle outliers.

To account for the fact that covered airports have two 18-month periods after which their AIR-21 coverage is re-determined, the dependent variable is constructed such that the effect of treatment is evaluated over the next accumulative 36-month period. Due to this feature, a number of further adjustments are made to the sample.

First, to avoid over-counting treatment, I track each airport starting with their initial

AIR-21 coverage after the legislation, and remove observations within 36 months of each subsequent coverage.11 Conversely, I check backwards to remove periods when an airport appears as non-treated according to boarding data where it subsequently becomes treated within the next 36 months.12 Finally, I track non-treated obser- vations and remove non-treated periods that occur within 36 months of a previous non-treated period.

Table 1.1 provides some descriptive statistics of firms and markets in the data.

The final sample includes 26 unique airline firms, all at the consolidated parent com- pany level, and 2,786 markets. Time-series averages of the quarterly cross-sectional mean, median, and standard deviation for a variety of airline and market character- istics are shown for 7 sub-periods as well as averaged over the full sample period.

It is first worth emphasizing that the large number of markets in the sample allows precise estimation and statistically powerful inference of the effect of relative-to-rival cash and multimarket contact on pricing at the firm-market level, despite the some- what small number of firms. It is also interesting to note that after the legislation of

AIR-21, I observe a sharp increase in cash holdings accompanied by higher leverage

11An airport treated as of 2002Q1 should not be considered “re-treated” or “un-treated” in, say, 2002Q2 or 2003Q1 even though the top-two airline concentration ratio might suggest so. 12An airport treated in 2003Q1 should not be considered non-treated in, say, 2002Q1 even though the top-two airline concentration ratio might suggest so.

22 Table 1.1: Airline and Market Summary Statistics This table summarizes airline and market characteristics, which are computed for 7 sub-periods as well as for the entire sample period. Panel A documents airline statistics. The first row shows time- series averages of the number of firms in each period. The remainder of the panel presents time-series averages of cross-sectional means, medians, and standard deviations of firm financial positions (cash, short-term debt, and long-term debt, all scaled by total assets) and firm size measured by total assets (in $ billions). Panel B documents market statistics. Time-series averages of the number of routes (markets) and cross-sectional means, medians, and standard deviations of quarterly average market prices charged by all airlines; by low cost carriers (LCCs); by legacy airlines, market-level number of passengers (in thousands), market-level LCC passenger shares, and market size in terms of total dollar revenues (in $ millions) are reported.

N 2001-02’ 2003-04’ 2005-06’ 2007-08’ 2009-10’ 2011-12’ 2013-14’ Average Panel A. Airline Statistics Firm N 26 21 18 18 19 17 15 14 17

Cash Mean 0.09 0.13 0.15 0.18 0.16 0.19 0.17 0.15 Median 0.08 0.10 0.13 0.17 0.15 0.21 0.17 0.14 Std. Dev (0.05) (0.07) (0.08) (0.07) (0.05) (0.06) (0.07) (0.06)

Short-Term Debt 0.03 0.04 0.04 0.04 0.03 0.05 0.04 0.04 0.02 0.03 0.03 0.04 0.03 0.04 0.04 0.03 (0.04) (0.02) (0.02) (0.02) (0.02) (0.02) (0.02) (0.03)

Long-Term Debt 0.22 0.32 0.35 0.30 0.28 0.30 0.24 0.29 0.24 0.34 0.39 0.34 0.29 0.30 0.23 0.30 (0.11) (0.10) (0.12) (0.14) (0.10) (0.08) (0.08) (0.10)

Size ($ billion) 12.76 15.22 14.51 14.61 15.85 17.09 22.45 16.07 10.38 10.38 10.92 13.84 15.36 14.61 18.08 13.37 (8.30) (9.76) (8.94) (8.55) (8.78) (11.80) (14.80) (10.13) Panel B. Market Statistics Route N 2786 1925 1831 1791 1739 1623 1670 1735 1759

Fares Mean 269.07 231.60 227.60 226.64 230.61 221.77 245.54 236.12 Median 255.45 218.14 221.13 222.67 224.93 217.11 241.01 228.63 Std. Dev (88.94) (68.91) (59.94) (59.03) (58.80) (56.95) (60.36) (64.70)

LCC Fares 197.09 184.10 181.81 179.42 187.39 183.34 201.24 187.77 196.88 184.18 183.36 178.92 185.58 183.64 202.08 187.81 (57.22) (49.76) (41.84) (41.58) (44.51) (41.99) (50.10) (46.71)

Legacy Fares 266.86 225.99 221.75 216.73 223.90 218.61 244.91 231.25 252.16 211.85 213.67 210.28 214.42 213.09 239.32 222.11 (100.63) (75.02) (66.53) (61.61) (63.56) (61.47) (67.71) (70.93)

Passengers 30.80 26.83 30.80 33.85 32.76 31.75 32.19 31.28 (thousands) 20.23 19.74 21.09 24.32 24.07 22.81 21.76 22.00 (28.46) (22.32) (27.58) (29.83) (29.21) (30.34) (31.92) (28.52)

LCC Share 0.20 0.23 0.26 0.31 0.37 0.41 0.40 0.31 0.09 0.13 0.15 0.25 0.34 0.39 0.38 0.25 (0.26) (0.27) (0.28) (0.30) (0.30) (0.32) (0.31) (0.29)

Market Size 7.39 5.57 6.08 6.55 6.58 6.21 6.96 6.48 ($ million) 5.08 4.08 4.58 4.81 4.89 4.40 4.86 4.67 (7.32) (4.76) (5.30) (5.95) (6.15) (6.36) (7.45) (6.18)

23 as seen in Panel A. For instance, mean cash reserves increase from 9% of assets in the

2001-2002 period to 15% in the 2005-2006 period, while the average long-term debt to asset ratio increases from 22% to 35% over the same period. Moreover, Panel B of

Table 1.1 reveals a decrease in prices following the enactment of AIR-21, coinciding with an increase in passenger boarding shares by low cost carriers (LLCs) but not accompanied by a sustained reduction in the number of passengers. Average fares fall from $269.07 in 2001-2002 to $227.6 in 2005-2006 and low cost carrier (LCC) passenger shares rise from 20% to 26%, whereas the average number of passengers remains flat around 31 thousand.

These patterns imply intensified competition in the industry which does not ap- pear to be driven by supposed demand shocks around the 9/11 terrorist attacks, and is associated with the accumulation of corporate cash holdings. I provide more ev- idence on the latter in Section 1.4, which is consistent with the notion that firms would rationally build up their cash holdings in response to increased competition if indeed financial war chests come in handy under such competitive circumstances.

Next, I describe how the main variables in the analysis are constructed based on the dataset.

1.3.2 Variables

The outcome variable of interest in this study is the change in firm pricing policy, which is the difference between firm-market level price growth over the next 36 months and the previous 36 months, skipping the previous and next year. It is computed as

13 the second difference of log prices, denoted by ∆∆36Log(Market Fares). Key among

13Snider and Williams(2015) also use price growth differentials to show the impact of AIR-21. In untabulated analysis, I confirm that results are qualitatively similar using price growth over the next 36 months skipping the first year, ∆36Log(Market Fares), as an alternative dependent variable.

24 the explanatory variables, AIR-21 treatment is assigned as follows. Each year, I begin by computing airport-level concentration ratios as the passenger boarding shares of the top-two airlines at each airport. Then I assign a treatment indicator variable for each market, denoted by AIR21, that equals 1 if the concentration ratio at the origin airport two years prior to the beginning of the present year (i.e. the forcing variable) exceeds 50% and equals 0 otherwise.14

Table 1.2 lists airlines and airports in the sample, and provides airport-level statis- tics related to AIR-21 treatment. For each airport, the time-series average, minimum, and maximum share of passengers boarded by the top-two airlines, as well as the air- port’s average share of total U.S. enplanement are shown in Panel B. Airport lists are shown separately for those covered by AIR-21 at some point, in which case the initial year of coverage is shown, and those never covered. An interesting takeaway is that there is not only substantial cross-sectional variation in top-two airline concentration across airports, but also sizable time-series variation within airports. For example, the average top-two airline passenger shares for ORD (Chicago) and LAX (Los An- geles) are widely apart at 63.34% and 31.27% respectively, but the track record for

ORD itself also ranges from as low as 49.72% to as high as 75.45%. This indicates considerable within and cross-variation in AIR-21 treatment even though most cov- ered airports appear to have become initially covered as soon as the legislation went into effect.

The ex-ante variables to explain the changes in pricing strategies after AIR-21

14It is plausible that AIR-21 treatment at the destination airport equally serves as a source of intensified competition in a market through shifts in access to gates. Snider and Williams(2015) implement a simple two-dimensional regression discontinuity design and show that this is the case. In untabulated analysis, I assign AIR-21 treatment based on the top-two airline concentration ratio at the destination airport instead of the origin, and find similar results.

25 Table 1.2: Airlines and Airports This table presents lists of legacy airlines and low cost carriers (Panel A), as well as a list and summary statistics of origin airports included in our sample (Panel B). For each airport, the time- series average, minimum, and maximum share of passengers boarded by the top-2 airlines, as well as the airports average share of total U.S. enplanement are shown. Airport lists and statistics are presented separately for airports that are covered by AIR-21 at some point within the sample period, in which case the initial year of coverage is shown, and those that are never covered.

Panel A. List of Airlines Legacy Airlines: AA, AQ, AS, CO, DL, HA, HP, JI, NW, TW, UA, US, YV, YX Low Cost Carriers: B6, F9, FL, G4, N7, NJ, NK, SY, TZ, U5, VX, WN

Panel B. Airport Statistics Concentration (Top 2, %) Relative to AIR21 Concentration (Top 2, %) Relative to AIR21 Airports Average Min Max US (%) Year Airports Average Min Max US (%) Year B-1. Covered Airports ABQ 67.24 63.60 71.34 0.46 2001 MIA 63.03 53.95 70.02 2.25 2001 ANC 59.67 51.57 71.03 0.34 2001 MKE 48.39 37.97 57.02 0.50 2001 ATL 78.66 73.41 82.30 5.76 2001 MSP 71.72 58.56 79.77 2.28 2001 AUS 59.34 39.83 61.94 0.56 2002 MSY 47.57 40.55 53.14 0.63 2001 BHM 73.13 73.13 73.13 0.26 2003 OAK 74.73 67.46 82.42 0.84 2001 BNA 65.10 62.42 69.89 0.68 2001 OGG 60.46 48.76 69.33 0.39 2001 BUR 77.05 70.72 84.12 0.35 2001 OKC 49.52 43.36 55.68 0.26 2011 BWI 64.43 55.50 72.29 1.45 2001 ONT 61.03 55.63 63.95 0.42 2001 CLE 56.27 49.03 61.68 0.74 2001 ORD 63.34 49.72 75.45 4.70 2001 CLT 76.24 69.27 87.04 2.06 2001 PBI 47.69 38.49 58.75 0.43 2001 CVG 75.18 51.87 93.70 1.12 2001 PHL 58.29 50.59 67.51 1.95 2001 DAL 97.94 94.77 99.84 0.53 2001 PHX 69.22 57.35 72.92 2.74 2001 DCA 41.03 33.16 51.68 1.12 2002 PIT 52.49 31.53 80.76 0.88 2001 DEN 60.21 45.90 72.58 3.02 2001 PVD 59.88 54.42 66.74 0.35 2001 DFW 81.00 73.14 85.33 3.82 2001 RNO 62.57 56.04 67.63 0.34 2001 DTW 71.06 61.02 77.84 2.30 2001 SAN 48.29 45.36 52.38 1.16 2012 ELP 77.28 74.08 80.56 0.26 2001 SAT 55.28 53.66 57.49 0.53 2001 EWR 63.59 58.95 69.91 2.30 2001 SDF 47.67 41.31 54.44 0.26 2001 HNL 49.27 43.90 60.56 1.35 2011 SFO 50.04 41.32 57.18 2.47 2001 HOU 91.26 88.51 93.54 0.66 2001 SJC 60.05 53.51 66.30 0.72 2001 IAD 50.77 45.62 57.51 1.45 2001 SJU 58.61 44.33 68.58 0.66 2001 IAH 82.42 77.85 86.11 2.53 2001 SLC 72.81 66.77 82.63 1.36 2001 JAX 46.99 41.10 54.89 0.38 2003 SMF 61.97 58.07 68.31 0.64 2001 LAS 50.63 47.67 54.13 2.71 2003 SNA 41.74 33.63 53.14 0.60 2014 MCI 49.83 44.19 55.14 0.80 2005 STL 68.49 54.81 86.33 1.31 2001 MDW 84.30 73.29 92.57 1.22 2001 TPA 43.82 39.80 51.14 1.17 2012 MEM 69.17 55.87 76.89 0.72 2001 TUL 52.76 51.59 53.93 0.25 2001 MHT 57.53 55.43 61.10 0.27 2004 TUS 48.52 42.93 53.83 0.26 2007

B-2. Non-Covered Airports BDL 43.60 39.27 49.25 0.44 LGA 38.70 30.76 44.50 1.67 BOS 33.50 27.23 39.15 1.80 MCO 35.89 32.49 42.17 2.19 BUF 39.11 30.73 46.70 0.33 OMA 41.09 37.62 45.05 0.28 CMH 34.35 28.49 40.27 0.46 ORF 38.97 34.43 41.28 0.26 FLL 34.21 27.51 40.75 1.36 PDX 37.78 35.50 39.53 0.93 IND 31.27 26.93 35.19 0.54 RDU 35.32 30.13 40.52 0.64 JFK 41.35 32.42 45.80 2.75 RSW 35.49 22.82 45.16 0.46 LAX 31.27 26.41 36.03 4.16 SEA 46.83 42.46 49.98 2.05

26 competition shocks are relative-to-rival cash holdings and multimarket contact. For

both of these variables, the first step is to identify rivalry. Competitors of firm i in

market r are firms j (where j 6= i) also operating in that market. Relative-to-rival

financial variables, cash and various control measures of debt, are then defined as the excess financial positions of a firm relative to the average of all rival firms. They are computed each quarter as follows,

n 1 Xr RelF in = F in − F in ir i n − 1 j r j=1

where j 6= i, nr denoting the number of firms in market r, and F inj denoting financial

variables scaled by total assets. To construct multimarket contact, I follow Evans

and Kessides(1994) and begin by counting the number of markets in which firm i

encounters rival company j, which can be written as

n X aij = DirDjr r=1

where n is the number of markets and Dir is a dummy variable equal to one if firm i

operates in market r and 0 otherwise. Then each quarter, firm i’s average multimarket

contact with rivals in market r is

n 1 Xr MC = a ir n − 1 ij r j=1

MCir essentially captures the average number of markets jointly contested by firm i and its rivals, and is logarithmized in the triple difference regressions. RelF inir and

MCir are both winsorized at the bottom and top 1% to lessen the impact of outliers, and are measured one year prior to AIR-21 treatment.15

Table 1.3 describes the quantities of the outcome variable (i.e. changes in firm

15A concern here is that not all rivals may be equally important. For example, in the first quarter of 2014, Virgin America had a 23.4% passenger share in the BOS (Boston)-SFO (San Francisco)

27 Table 1.3: Key Variable Statistics This table presents summary statistics of airline price growth differentials, relative-to-rival finances, and multimarket contact over the sample period 2001Q1 to 2014Q4. Price growth differential is defined as the difference between the next 36-month’s and previous 36-month’s price growth, skip- ping the previous year and next year, and is computed by taking the second difference of log prices (∆∆36Log(Market Fares)). Rivals are identified as firms operating in market r. Relative-to-rival finances and multimarket contact are computed following equations (1) and (2) as detailed in Sec- tion 1.3. 1 Pnr (1) RelF inir = F ini − F inj where j 6= i nr −1 j=1 1 Pnr Pn (2) MCir = aij where aij = DirDjr and j 6= i nr −1 j=1 r=1 Price growth differentials are computed each quarter for each firm-market separately for nonstop and connecting flights. Relative-to-rival finances and multimarket contact variables are computed each quarter for each firm-market. Then, time-series averages of the cross-sectional mean, standard deviation, minimum, median, and maximum for each of these variables are presented in Panel A. Price growth differentials are presented as percentage growth rates, financial variables as percent- ages of total assets, and multimarket contact as average number of routes. The number of firms and markets are shown as well. Panel B presents Pearson correlations between the relative-to-rival finance and multimarket contact variables. Panel A. Average Cross-Sectional Statistics Mean Std. Dev Min Median Max Price Growth Differential 4.28 33.72 -350.95 3.91 350.66

Relative-to-Rival Finance Cash -0.03 8.89 -28.37 -0.42 29.44 Short-Term Debt -0.20 3.53 -13.76 -0.35 13.92 Long-Term Debt -0.64 13.53 -41.89 -0.04 41.07 Total Debt -0.85 15.09 -48.51 -0.33 47.92

Multimarket Contact 193.18 103.47 10.70 191.28 422.00

No. of Firms 26 No. of Markets 2786

Panel B. Pearson Correlations Relative-to-Rival Finance Multimarket Cash ST Debt LT Debt T Debt Contact Relative-to-Rival Finance Cash 1 0.08 -0.05 -0.02 -0.09 Short-Term Debt 0.08 1 0.24 0.48 -0.05 Long-Term Debt -0.05 0.24 1 0.96 -0.04 Total Debt -0.02 0.48 0.96 1 -0.05 Multimarket Contact -0.09 -0.05 -0.04 -0.05 1

28 pricing policy) and the ex-ante explanatory variables explained above (i.e. relative-to- rival finances and multimarket contact). In Panel A, time-series averages of the cross- sectional mean, standard deviation, minimum, median, and maximum are shown for each variable. The mean difference between price growth over the next and previous

36 months is 4.28% with a standard deviation of 33.72%. There is substantial cross- sectional variation in both relative-to-rival cash holdings and multimarket contact, mitigating concerns that there is unlikely much action on these dimensions. Relative- to-rival cash to asset ratios range from -28.37% to 29.44% with an average standard deviation of 8.89%. The average multimarket contact is 193.18 routes, ranging from

10.7 to as large as 422, with a standard deviation of 103.47.16 Panel B presents Pear- son correlations between relative-to-rival finances and multimarket contact. None of the financial variables are highly correlated with multimarket contact. The corre- lation between relative-to-rival cash and multimarket contact is -0.09, leaving little room to suspect that one strongly influences the other.

1.3.3 Empirical Specification

To explore how cash holdings affect firm pricing behavior, and whether multimar- ket contact strengthens or weakens this effect, I employ a triple difference framework on firm-market and quarter panel data where I regress price growth differentials on market. Between its rival firms American Airlines and , each of which had 2.4% and 41.7% passenger shares respectively, United Airlines would have been a more formidable and important competitor for Virgin America in the BOS-SFO market. In light of such situations, equally weighting all rival firms in the same market may incorrectly assign a firm with more or less multimarket contact or relative-to-rival cash than is actually relevant. To take this into account, I reconstruct these variables weighting rival firms in each market by their passenger boardings. Doing so essentially leaves the results of this paper unchanged. These alternative results are reported in the Appendix (Table A.1). 16To facilitate the interpretation of economic magnitudes of the regression results presented in Section 1.4, I also report the key variable statistics for a number of select subsamples (the results for which economic magnitudes are discussed in the paper) in the Appendix (Table A.2).

29 relative-to-rival cash, multimarket contact, and an AIR-21 treatment indicator, as well as a host of control variables and fixed effects. To capture a plausibly exogenous local treatment effect, this regression is run on progressively narrower windows sur- rounding the 50% top-two airline passenger share threshold for AIR-21 treatment in market origin airports. The baseline model is specified as follows.

∆∆36log (Pirt) = β0 + β1 · MCirt−1 + β2 · RelCashirt−1 + β3 · AIR21rt

+ β4 · (MCirt−1 × RelCashirt−1)

+ β5 · (AIR21rt × MCirt−1)

+ β6 · (AIR21rt × RelCashirt−1)

+ β7 · (AIR21rt × MCirt−1 × RelCashirt−1)

0 + γ · Xirt−1 + ai + br + ct + irt

In the baseline specification as well as in additional tests, I control for relative-to- rival short-term debt, long-term debt, and their cross interactions with multimarket contact and AIR-21 treatment to tease out the impact of cash holdings distinct from well-established leverage effects.17 I further include as control variables an indicator for whether the service is a direct flight, firm’s passenger share in the market, one year lagged firm size measured by log assets, number of rivals in the market, passenger share of low cost carriers (LCCs) in the market, and market size in terms of log dollar revenues. Firm, market, and time fixed effects are also included in all specifications.18

17In their concluding remark, Opler, Pinkowitz, Stulz, and Williamson(1999) emphasize the fact that determinants of cash holdings and debt are closely related. In an examination of this point, Acharya, Almeida, and Campello(2007) show that there are instances where cash is not necessarily negative debt, indicating that cash should have implications that are distinct from those of debt. Fresard(2010) shows this in the context of product market outcomes. 18To address concerns of multicollinearity between variables included in the regressions, I report a correlation table (Figure A.1) in the Appendix. While there is some correlation among certain variables, they are moderate in magnitude and intuitive to understand. For example, the highest correlation is 0.59 between firm size and multimarket contact (MC), which is not surprising since

30 To understand the meaning of the coefficients, recall that there are two hypotheses

being tested. First, I examine the implications of ‘deep pocket’ theories about the

mechanism through which cash may affect product market outcomes, namely that

cash holdings provide the financial flexibility that enables firms to price aggressively

(see Bolton and Scharfstein(1990)). To do this, I test whether firms with large ex-ante

relative-to-rival cash reserves reduce prices more aggressively after AIR-21 competi-

tion shocks. This test is captured by β6, the coefficient on the interaction term of

AIR21 and RelCash, for which the hypothesized prior is a negative value. Second, incorporating insights from the industrial organization literature on multimarket con- tact and collusive behavior (see Bernheim and Whinston(1990)), I test whether high multimarket contact attenuates the competitive effect of cash. The attenuation effect of multimarket contact is captured by β7, which is expected to take the opposite sign

of β6 and therefore have a positive value.

In the following section, I present results from baseline regressions, a number of subsample analyses, and additional robustness tests to cement the argument of the paper.

1.4 Results

1.4.1 Main Results

Table 1.4 presents results from the baseline firm-market-quarter level triple dif- ference regressions of price growth differentials on AIR-21 treatment, relative-to-rival cash holdings, multimarket contact, and their cross interactions. Cash and multimar- ket contact are both measured one year prior to AIR-21 treatment. Firm, market,

larger firms will tend to operate in a greater number of major routes where they will make contact with a broader set of rivals.

31 and time fixed effects are controlled for, and standard errors adjusted for clustering at the firm and market levels are reported in parentheses.19 The five columns show the main results, where triple difference regressions are run using progressively narrower sample windows surrounding the 50% AIR-21 treatment cutoff.

Moving away from the full sample, results imply a strong causal impact of cash on changes in firm pricing policy which is attenuated by higher levels of multimarket contact. In the sample of observations where the value of the forcing variable, the ex-ante top-two airline concentration ratio at the origin airport of the market, ranges from 35% to 65% (i.e. 15% above and below the 50% cutoff), a one standard devia- tion (i.e. 8.55 percentage point) increase in ex-ante relative-to-rival cash holdings as a fraction of assets leads to 15.22 percentage points lower price growth over the next 36 months compared to the previous 36 months (i.e. 0.47 standard deviations lower price growth differential). This competitive effect of cash in which cash-rich firms respond to competition shocks by pricing more aggressively, captured by the coefficient on the interaction term of AIR-21 treatment and relative-to-rival cash, is rapidly eroded by increases in multimarket contact: a 5.3% increase completely cancels out the effect.

This is consistent with the hypothesis that firms facing higher chances of retaliation from rivals are less likely to make use of their financial war chests. These effects are

19A stricter test is to zero in on the change in pricing behavior of the same firm in the same market after being treated by AIR-21, which can be implemented by controlling for firm-by-market fixed effects instead of firm and market fixed effects separately. Results from this alternative specification, which are robust if not stronger, are reported in Panel A of Table A.3 in the Appendix. It is also important to verify the robustness of the results to adjusting for alternative forms of error clustering. For example, a hypothetical measurement error in a firm-level financial variable in a particular time period will likely cause errors to correlate across all market observations for the same firm at that time. Panel B of Table A.3 shows robust results after adjusting standard errors for such a case of firm-time level clustering, as well as a number of other error clustering schemes.

32 Table 1.4: Cash, Competition, and Multimarket Contact This table presents results from firm-market-quarter level triple-difference regressions of 12-quarter price growth differentials on an AIR-21 coverage indicator (AIR21), ex-ante relative-to-rival cash holdings (Cash), ex-ante multimarket contact (MC), and their interactions. Specifically, the depen- dent variable (∆∆36Log(Market Fares)) is defined as the difference between the next 36-month’s and previous 36-month’s price growth, skipping the previous year and next year, and is computed by taking the second difference of log prices. The AIR-21 dummy variable equals to 1 if the top-2 airlines’ passenger boardings exceed 50% of the origin airport total during the calendar year 2 years prior to the current quarter, and equal to 0 otherwise (In untabulated analysis, I find similar re- sults using an AIR-21 treatment indicator for the destination airport). Relative-to-rival cash and multimarket contact are constructed as described in Section 1.3, and are measured as of the pre- vious year. Other control variables include relative-to-rival short-term debt, long-term debt, their interactions with multimarket contact and the AIR-21 indicator, a nonstop service indicator, firm’s passenger share in the market, firm size measured by the logarithm of total assets, number of rivals in the market, total passenger share of low cost carriers (LCCs) in the market, and market size measured as the logarithm of market-level aggregate dollar revenues. The five columns show results from triple difference regressions using progressively narrower sample windows surrounding the 50% AIR-21 treatment cutoff: (1) full sample, (2) 15% above and below the 50% cutoff, (3) 10% above and below the 50% cutoff, (4) 5% above and below the 50% cutoff, and (5) 2.5% above and below the 50% cutoff. Firm, market, and time fixed effects are included in all specifications. Standard errors adjusted for clustering at firm and market levels are reported in parentheses (*** p<0.01, ** p<0.05, * p<0.1).

Variables Dependent Variable: ∆∆36Log(Market Fares) Full Sample 15% 10% 5% 2.50%

AIR21 × Cash -0.518 -1.776*** -1.942*** -2.967*** -3.628** (0.543) (0.488) (0.665) (0.770) (1.706)

AIR21 × MC × Cash 0.092 0.333*** 0.348** 0.590*** 0.713* (0.110) (0.103) (0.132) (0.159) (0.382)

MC -0.038*** -0.033** -0.027 -0.023 -0.007 (0.012) (0.013) (0.020) (0.030) (0.034)

Cash 1.193*** 1.619*** 1.306** 2.109*** 2.672** (0.395) (0.465) (0.493) (0.670) (1.216)

MC × Cash -0.272*** -0.360*** -0.321*** -0.485*** -0.610* (0.082) (0.096) (0.102) (0.138) (0.299)

AIR21 0.043 0.082 0.112 -0.245* -0.326** (0.070) (0.089) (0.103) (0.124) (0.151)

AIR21 × MC -0.018 -0.026 -0.030 0.040 0.040 (0.013) (0.017) (0.020) (0.025) (0.033)

Observations 15,958 9,097 6,047 2,828 1,701 Firms 19 19 19 19 18 Markets 1,941 1,404 1,036 561 394 Firm / Market / Time FE Yes Yes Yes Yes Yes Controls Yes Yes Yes Yes Yes R-squared 0.226 0.271 0.301 0.349 0.401

33 not only robust, but even stronger as I move to narrower samples where the regres- sions are focused on observations with top-two airline concentration ratios 10%, 5%, and 2.5% above and below the 50% threshold. For example, in the second-narrowest sample 5% above and below the cutoff, a one standard deviation (i.e. 8.93 percentage point) increase in the relative-to-rival cash to asset ratio results in a reduction of price growth by 26.52 percentage points (i.e. 0.87 standard deviations lower price growth differential), which is wiped out by 5% higher multimarket contact.20 This is not surprising since I am more likely to capture the action of interest with a sample closer to the treatment threshold, where observations are more likely to be similar across controlled or unobserved dimensions, if the treatment is truly discontinuous at that point. Similarly, it is also natural that the results are less significant with the full sample as it includes observations at the far ends of the forcing variable’s domain that are more likely to be heterogeneous across unobservable dimensions. For exam- ple, observations where the top-two airline concentration ratio exceeds 70% exhibit substantially higher price growth differentials and lower low cost carrier passenger shares compared to observations below the 50% cutoff, indicating that differences in

20As an observational example to help understand the results, consider American Airlines and two of its markets, San Diego (SAN) to New York (JFK) and San Diego (SAN) to Chicago (ORD). At the end of 2010, American held 22% of its assets in cash. At the end of 2009, the two largest airlines in terms of passenger boardings, United and Southwest, boarded 49% of all passengers at SAN. At the end of 2010, the two largest airlines, Delta and Southwest, crossed the 50% AIR-21 threshold at SAN by boarding 50.3% of the airport’s passengers. As a result, SAN was covered by AIR-21 in 2012. American subsequently reduced its price growth by 8.2 percentage points for its service from SAN to JFK, where its rival, jetBlue, held 15% of its assets in cash as of 2010 year-end, considerably less than American. On the other hand, American raised its price growth rate by 12.4 percentage points on its route SAN to ORD, where its rival, United, held 24% of its assets in cash, even more than the amount American held. American also happened to overlap with United across a larger number of routes.

34 market dynamics may be large enough to dominate the impact of AIR-21.21

It is worth noting that the coefficients on relative-to-rival cash and its interac- tion with multimarket contact imply that firms with more cash price aggressively even absent AIR-21 shocks. For example, the average log multimarket contact in the

15% window sample is 5.3, at which the coefficient for relative-to-rival cash when

AIR21=0 is 1.619 + (−0.360) × 5.3 = −0.289. In a simple OLS regression of price growth differentials on relative-to-rival cash and control variables, the coefficient on relative-to-rival cash is -0.252 with an adjusted standard error of 0.101.

Interestingly, I do not find significant coefficients on the interaction terms from difference-in-differences analyses where I regress changes in firm pricing policy on

AIR-21 treatment and each of relative-to-rival cash or multimarket contact. I in- terpret this as an indication that cash reserves or multimarket contact alone do not suffice to influence firm pricing policy to a great extent, but rather it is the inter- play of financial and strategic rivalry considerations that is of critical importance. In addition, using 18 month horizons instead of 36 months to measure changes in firm pricing policy yields no results in my analysis, which is consistent with the idea that airports are likely to fully utilize the two 18 month periods given to them before their

AIR-21 coverage is re-determined to address the problems raised by the FAA. Their

21In the Appendix (Table A.4), I report the means of numerous observable variables for AIR- 21 control (AIR21=0) and treatment (AIR21=1) observation groups, as well as p-values from t- tests of their differences. Comparison of the means of firm-level and airport-level variables suggest that airline companies and market origin airports are similar across control and treatment groups, although airport characteristics are marginally different as the sample is broadened away from the tighter window around the 50% top-two airline passenger share threshold. At market and firm- market levels, there are more statistically significant differences across a number of variables, but those differences are small in magnitude in the full sample and further diminish as the sample window is narrowed to 15% above and below the 50% cutoff. Beyond explicitly controlling for a large number of these variables to the extent that multicollinearity issues do not arise, I also include firm, market, and time fixed effects, and adjust standard errors for clustering at firm and market levels to address such differences.

35 tendency to postpone the remedy until the last minute may arise from a variety of reasons, such as relationships with incumbent airlines, physical capacity constraints, or distracted management.

It is important to realize that the main results imply a cash impact on firm com- petition strategy that is markedly different from previously documented effects of debt. Busse(2002) finds that airline companies with greater leverage are more likely to start price wars, consistent with the notion that risk-shifting by highly indebted

firms leads them to ‘gamble harder’ by lowering prices (see Brander and Lewis(1986),

Maksimovic(1988)). Thus if cash were simply the flip side of debt, one would expect airlines with smaller cash holdings to price more aggressively following a competition shock, which is the opposite from what I find. While it should be kept in mind that the relationship between capital structure and product market competition is likely to vary across industry structures (see Phillips(1995)), the airline industry in particular highlights the distinct effects of cash and debt since larger cash holdings and higher leverage both appear to result in more aggressive pricing behavior, contradicting the notion that cash is merely negative debt.

In the following subsections, I supplement the main results with a variety of alter- native tests to cement the story that cash provides financial flexibility which enables

firms to compete aggressively, and that potential rival retaliation attenuates the com- petitive benefit of this flexibility.

1.4.2 Subsample Analyses

While the baseline results imply a competitive benefit of cash that varies with mul- timarket contact on average, an important aspect of the AIR-21 regulation is that

36 it likely affects heterogeneous firms in the same market differentially. The reason is that AIR-21 pushes covered airports to redirect gate allocations toward new entrant airlines, mostly low cost carriers (LCCs), and incumbent airlines with small market shares, as documented by Snider and Williams(2015). It is therefore important to verify these heterogeneous effects in subsample analyses to clarify the source of the main results.

Table 1.5 shows triple difference regression results from splitting the sample of

firms into LCCs and legacy airlines. A natural way to think about how LCCs and legacy airlines would respond to AIR-21 treatment is to view them as serving loosely substitutable, but distinct, clienteles. Because AIR-21 tends to open up airports to new entrant LCCs, incumbent LCCs serving the same clientele are likely to re- spond aggressively across the board, whereas legacy airlines will tend to respond less unanimously given their services are to some extent differentiated from LCCs’ al- though still quite substitutable. Consistent with this hypothesis, I find that LCCs respond to AIR-21 shocks by pricing more aggressively irrespective of their cash hold- ings, while legacy airlines do so conditional on holding large cash reserves. Both of their responses, however, are dampened by higher multimarket contact. While per- forming each of the subsample analyses on increasingly closer windows around the

50% treatment threshold presents possible small-sample issues (e.g. the number of

firm-market-quarter observations drop below 1,000 for the LCC subsample when the forcing variable range is within 5% above and below the 50% cutoff), the results are strikingly robust and consistent for legacy airlines as I close in from the full sample.

In response to AIR-21 competition shocks, legacy airlines with one standard devia- tion (i.e. 8.42 percentage points) greater cash holdings as a fraction of assets relative

37 Table 1.5: Low Cost Carriers vs. Legacy Airlines

This table presents subsample results from triple-difference regressions of 12-quarter price growth differentials (∆∆36Log(Market Fares)) on an AIR-21 coverage indicator (AIR21), ex-ante relative-to-rival cash holdings (Cash), ex-ante multimarket contact (MC), and their interactions. Two sets of regressions are run for subsamples restricted to include only (1) low cost carriers and (2) legacy airlines. In each subsample analysis, the results are shown in five columns where the triple difference regression is run using progressively narrower sample windows surrounding the 50% AIR-21 treatment cutoff: (1) full sample, (2) 15% above and below the 50% cutoff, (3) 10% above and below the 50% cutoff, (4) 5% above and below the 50% cutoff, and (5) 2.5% above and below the 50% cutoff. Firm, market, and time fixed effects as well as control variables are included in all specifications as in previous analysis. Standard errors adjusted for clustering at firm and market levels are reported in parentheses (*** p<0.01, ** p<0.05, * p<0.1).

Variables Dependent Variable: ∆∆36Log(Market Fares) Low Cost Carriers Legacy Airlines Full Sample 15% 10% 5% 2.50% Full Sample 15% 10% 5% 2.50%

AIR21 × Cash -0.033 -1.133 -0.057 -0.401 5.279*** -0.858 -2.389*** -2.437*** -4.371*** -7.341*** (0.609) (0.638) (0.759) (0.982) (1.018) (0.695) (0.512) (0.779) (0.857) (2.110) 38 AIR21 × MC × Cash -0.015 0.162 -0.113 -0.037 -1.156*** 0.128 0.424*** 0.411** 0.892*** 1.499*** (0.115) (0.120) (0.155) (0.217) (0.258) (0.146) (0.123) (0.147) (0.168) (0.470)

MC -0.070** -0.098*** -0.120*** -0.142** -0.398** -0.040** -0.045* -0.021 0.010 0.022 (0.023) (0.019) (0.021) (0.035) (0.152) (0.015) (0.023) (0.028) (0.034) (0.050)

Cash 0.402 0.936 0.319 -0.072 -4.581*** 1.809*** 2.188*** 1.702* 2.967** 3.984 (0.798) (0.810) (0.995) (1.297) (0.962) (0.366) (0.570) (0.819) (1.029) (2.531)

MC × Cash -0.079 -0.151 -0.016 0.119 0.871** -0.352*** -0.437*** -0.403** -0.722*** -0.953 (0.147) (0.136) (0.210) (0.304) (0.247) (0.090) (0.118) (0.158) (0.204) (0.561)

AIR21 -0.226** -0.334*** -0.456*** -0.321 -2.313** 0.054 0.058 0.104 -0.266 -0.218 (0.090) (0.073) (0.094) (0.207) (0.591) (0.095) (0.124) (0.140) (0.164) (0.233)

AIR21 × MC 0.029 0.058*** 0.091*** 0.053 0.494** -0.021 -0.023 -0.030 0.041 0.003 (0.016) (0.012) (0.019) (0.037) (0.147) (0.018) (0.024) (0.028) (0.032) (0.050)

Observations 4,623 2,655 1,773 893 566 11,086 6,219 4,075 1,845 1,061 Firms 7 7 7 6 6 12 12 12 12 11 Markets 961 675 470 256 194 1,625 1,140 818 443 294 Firm / Market / Time FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Controls Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes R-squared 0.300 0.373 0.411 0.459 0.573 0.245 0.293 0.327 0.378 0.432 to rivals exhibit as much as 61.8 percentage points lower price growth over the next

36 months compared to the previous 36 months (i.e. 1.71 standard deviations lower price growth differential), and a mere 4.9% increase in multimarket contact cancels out this effect. The legacy airline subsample results have important ramifications.

Not only are legacy airlines such as Delta, American, and United representative firms of the airline industry, but the manner in which they respond to AIR-21 competition shocks is also representative of how most firms in other industries with differentiated products and incompletely substitutable clienteles are likely to behave.

In Table 1.6, I present results from subsample analyses for firms that are clear market dominators ex-ante (i.e. firms with passenger shares above 45% one year prior to treatment) and those that are not (i.e. firms with passenger shares below

45% one year prior to treatment). Given that the aim and result of AIR-21 are to reallocate gates toward new entrants and dominated incumbents at covered airports, one would expect to see the effect of AIR-21 as a competition shock for firms that have high market shares to begin with, but not for ex-ante market laggards for whom treatment would rather come as a blessing. This prediction is confirmed in the re- sults. I do not find that firms with low ex-ante market shares respond to competition shocks in terms of their pricing strategies, irrespective of their cash holdings or multi- market contact. On the other hand, firms that dominate the market prior to AIR-21 treatment respond by pricing more aggressively with their cash holdings (e.g. in the window sample 15% above and below the 50% cutoff, a one standard deviation, or

8.49 percentage point, increase in relative-to-rival cash to asset ratio corresponds to

22.41 percentage points lower price growth, or 0.76 standard deviations lower price growth differential), and greater ex-ante multimarket contact erodes this effect (e.g.

39 Table 1.6: Ex-Ante Market Dominators vs. Laggards

This table presents subsample results from triple-difference regressions of 12-quarter price growth differentials (∆∆36Log(Market Fares)) on an AIR-21 coverage indicator (AIR21), ex-ante relative-to-rival cash holdings (Cash), ex-ante multimarket contact (MC), and their interactions. Two sets of regressions are run for subsamples restricted to include only (1) ex-ante market dominators and (2) ex-ante market laggards. Ex-ante market dominators (laggards) are defined as firms with ex-ante, i.e. measured 1 year prior to a given point of time, passenger shares equal to or larger than (less than) 45% in a given market and quarter. In each subsample analysis, the results are shown in five columns where the triple difference regression is run using progressively narrower sample windows surrounding the 50% AIR-21 treatment cutoff: (1) full sample, (2) 15% above and below the 50% cutoff, (3) 10% above and below the 50% cutoff, (4) 5% above and below the 50% cutoff, and (5) 2.5% above and below the 50% cutoff. Firm, market, and time fixed effects as well as control variables are included in all specifications as in previous analysis. Standard errors adjusted for clustering at firm and market levels are reported in parentheses (*** p<0.01, ** p<0.05, * p<0.1).

Variables Dependent Variable: ∆∆36Log(Market Fares) Dominators (Market Share ≥ 45%) Laggards (Market Share < 45%) Full Sample 15% 10% 5% 2.50% Full Sample 15% 10% 5% 2.50%

AIR21 × Cash -1.320 -2.643** -3.062** -5.437** -11.549 0.737 -0.467 0.163 -1.060 -1.541 (0.866) (0.939) (1.164) (1.979) (8.397) (0.894) (1.259) (1.389) (1.614) (2.696) 40

AIR21 × MC × Cash 0.280 0.477** 0.515** 1.091** 2.247 -0.188 0.083 -0.079 0.186 0.255 (0.165) (0.197) (0.233) (0.414) (1.791) (0.194) (0.259) (0.281) (0.311) (0.531)

MC -0.050** -0.040 -0.023 -0.028 0.130 -0.051*** -0.070*** -0.076** -0.002 -0.013 (0.023) (0.026) (0.035) (0.064) (0.113) (0.013) (0.019) (0.027) (0.050) (0.070)

Cash 2.653*** 2.362*** 1.599 2.349 7.341 -0.698 0.051 -1.064 0.415 -0.740 (0.627) (0.651) (0.966) (1.338) (7.024) (0.772) (0.994) (1.048) (1.061) (1.964)

MC × Cash -0.563*** -0.464*** -0.344* -0.537* -1.584 0.078 -0.107 0.117 -0.173 0.039 (0.098) (0.122) (0.189) (0.283) (1.544) (0.166) (0.201) (0.215) (0.219) (0.402)

AIR21 0.009 0.076 0.036 -0.072 0.845 0.002 -0.061 -0.067 -0.264 -0.378 (0.130) (0.174) (0.183) (0.184) (0.550) (0.117) (0.148) (0.208) (0.191) (0.244)

AIR21 × MC -0.012 -0.032 -0.027 -0.001 -0.226* -0.007 0.008 0.012 0.044 0.087 (0.023) (0.032) (0.032) (0.036) (0.113) (0.022) (0.029) (0.042) (0.035) (0.057)

Observations 7,940 4,407 2,887 1,331 759 7,615 4,318 2,858 1,325 827 Firms 18 17 16 14 14 18 18 18 18 16 Markets 1,696 1,128 795 427 277 1,270 866 603 311 221 Firm / Market / Time FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Controls Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes R-squared 0.301 0.349 0.383 0.423 0.469 0.290 0.340 0.357 0.413 0.466 the reduction in price growth associated with larger cash reserves is wiped out by

5.5% higher multimarket contact).

From the subsample analyses above, I conclude that the average effects shown in

the main results are driven by legacy airlines and ex-ante market dominators, which

are the representative firms for which AIR-21 serves as a valid competition shock.

1.4.3 Do Firms Hold Cash Anticipating AIR-21 Coverage?

To further cement the main results of this paper, I conduct a number of robustness

checks to address some remaining concerns. One concern is that firm cash holdings

may still be endogenously linked to ex-post changes in pricing policy since firms may

be able to predict AIR-21 coverage and accumulate cash reserves in advance. If this

is the case, the causal interpretation of cash on pricing fails because firms that are

intent on pricing aggressively turn out to be the ones that hoard cash.

In Table 1.7, I address this particular issue by using relative-to-rival cash holdings

measured 3 or 4 years prior to AIR-21 treatment instead of one year in the triple

difference regressions. This alleviates the remaining endogeneity concern since firms

are unlikely to be able to predict AIR-21 coverage of an originating airport in a

market so far in advance. To do so, they would have to be able to foresee airport

total passenger boardings as well as boardings by the two largest airlines 1 or 2

years in advance, which is implausible. Panel A and Panel B report coefficients on

the AIR21×Cash and AIR21×MC×Cash interaction terms from the triple difference regressions each using 3-year and 4-year lagged relative-to-rival cash as the Cash variable. My results are strongly robust to these alternative specifications across all forcing variable windows, even in the full sample when I use 4-year lagged cash.

41 Table 1.7: Cash 3 to 4 Years Prior, Competition, and Multimarket Contact

This table presents results from triple-difference regressions of 12-quarter price growth differentials (∆∆36Log(Market Fares)) on an AIR-21 coverage indicator (AIR21), ex-ante multimarket contact (MC, measured as of the previous year), ex-ante relative-to-rival cash holdings (Cash 3Y, measured 3 years prior, or Cash 4Y, measured 4 years prior), and their interactions. The results are shown in five columns where the triple difference regression is run using progressively narrower sample windows surrounding the 50% AIR-21 treatment cutoff: (1) full sample, (2) 15% above and below the 50% cutoff, (3) 10% above and below the 50% cutoff, (4) 5% above and below the 50% cutoff, and (5) 2.5% above and below the 50% cutoff. Panel A reports results from using relative-to-rival cash measured 3 years prior, and Panel B reports results from using relative-to-rival cash measured 4 years prior. Firm, market, and time fixed effects as well as control variables are included in all specifications as in previous analysis. Standard errors adjusted for clustering at firm and market levels are reported in parentheses (*** p<0.01, ** p<0.05, * p<0.1).

Variables Dependent Variable: ∆∆36Log(Market Fares) Panel A. Relative-to-Rival Cash Lagged by 3 Years Full Sample 15% 10% 5% 2.50%

AIR21 × Cash 3Y -0.370 -1.303** -2.194** -3.208** -3.244** (0.448) (0.602) (0.818) (1.381) (1.146)

AIR21 × MC × Cash 3Y 0.092 0.244* 0.444** 0.643** 0.591** (0.100) (0.125) (0.172) (0.272) (0.235)

Observations 10,528 5,904 3,829 2,058 1,275 R-squared 0.284 0.321 0.346 0.352 0.434

Panel B. Relative-to-Rival Cash Lagged by 4 Years Full Sample 15% 10% 5% 2.50% AIR21 × Cash 4Y -1.160** -2.293*** -3.240*** -5.138*** -5.101** (0.476) (0.476) (0.946) (1.207) (2.199)

AIR21 × MC × Cash 4Y 0.236** 0.430*** 0.629*** 0.988*** 0.961** (0.101) (0.098) (0.201) (0.252) (0.432)

Observations 9,861 5,413 3,448 1,782 1,049 R-squared 0.282 0.328 0.349 0.367 0.464

Firm / Market / Time FE Yes Yes Yes Yes Yes Controls Yes Yes Yes Yes Yes

Focusing on a narrower sample where the forcing variable ranges from 45% to 55%

(i.e. 5% above and below the treatment threshold), a one standard deviation (i.e.

42 7.62 percentage point) increase in relative-to-rival cash holdings measured 4 years prior to treatment leads to 39.17 percentage points less price growth in the next 36 months compared to the previous 36 months (i.e. 1.29 standard deviations lower price growth differential). The competitive effect of cash is cancelled out by a 5.2% increase in multimarket contact. Even still, one might argue that if boardings are persistent,

firms would still be able to predict AIR-21 coverage. While this may be true for regions in the sample where the forcing variable takes values far away from the 50% treatment threshold, the quasi-RD approach of my analysis sidesteps this problem as

firms would not be able to predict whether the top-two airline concentration ratio of an airport will fall slightly above or below the 50% cutoff.

1.4.4 Is the AIR-21 Discontinuity Spurious?

Another concern is that the triple difference regression results may not be con-

fined to the case where the treatment cutoff of the forcing variable is 50%. Suppose that there is some other force at play apart from AIR-21 competition shocks that influences firm pricing strategies, and that this force is present not only when the forcing variable equals 50% but in other regions of the sample as well. Then, it might be the case that the AIR-21 indicator in the regressions captures an effect that is not due to AIR-21, purely by chance. To tackle this problem, I run placebo tests using alternative and arbitrary values as the treatment cutoff.

Table 1.8 shows results from these placebo tests. I run the same triple difference regressions, but with a fake AIR-21 indicator variable that equals one if the top-two airline concentration ratio of the originating airport of the market is greater than an

43 Table 1.8: Placebo Tests with Alternative Treatment Thresholds

This table presents placebo test results from triple-difference regressions of 12-quarter price growth differentials (∆∆36Log(Market Fares)) on an AIR-21 coverage indicator (AIR21), ex-ante relative-to-rival cash holdings (Cash), ex-ante multimarket contact (MC), and their interactions. Two sets of placebo regressions are run using alternative top-2 airline passenger share thresholds, 40% and 60%. In each placebo analysis, the results are shown in five columns where the triple difference regression is run using progressively narrower sample windows surrounding the 50% AIR-21 treatment cutoff: (1) full sample, (2) 15% above and below the 50% cutoff, (3) 10% above and below the 50% cutoff, (4) 5% above and below the 50% cutoff, and (5) 2.5% above and below the 50% cutoff. Firm, market, and time fixed effects as well as control variables are included in all specifications as in previous analysis. Standard errors adjusted for clustering at firm and market levels are reported in parentheses (*** p<0.01, ** p<0.05, * p<0.1).

Variables Dependent Variable: ∆∆36Log(Market Fares) Cutoff: Top-2 Airline Passenger Share = 40% Cutoff: Top-2 Airline Passenger Share = 60% Full Sample 15% 10% 5% 2.50% Full Sample 15% 10% 5% 2.50%

AIR21 × Cash 0.444 0.520 0.843 -0.211 1.780 0.568 0.591 1.271 0.851 -2.408 (0.504) (0.494) (0.530) (0.767) (2.744) (0.599) (0.593) (0.908) (1.425) (3.480) 44 AIR21 × MC × Cash -0.093 -0.100 -0.161 0.038 -0.390 -0.096 -0.068 -0.185 -0.049 0.655 (0.105) (0.100) (0.111) (0.160) (0.555) (0.119) (0.120) (0.191) (0.282) (0.719)

MC -0.037*** -0.037** -0.046** -0.031 -0.057 -0.047*** -0.059** -0.068** -0.073** -0.113* (0.011) (0.013) (0.017) (0.024) (0.051) (0.012) (0.023) (0.025) (0.028) (0.059)

Cash 0.567 0.540 1.031* 1.855* 1.556 0.675 -0.033 -0.547 -0.180 2.210 (0.466) (0.484) (0.546) (0.958) (1.381) (0.425) (0.559) (0.756) (1.109) (1.739)

MC × Cash -0.150 -0.147 -0.234** -0.417** -0.268 -0.184** -0.066 0.014 -0.069 -0.513 (0.097) (0.097) (0.106) (0.194) (0.308) (0.085) (0.110) (0.149) (0.212) (0.368)

AIR21 0.108* -0.015 -0.005 0.077 -0.108 0.031 -0.061 -0.104 -0.181 -0.397 (0.060) (0.078) (0.089) (0.109) (0.241) (0.062) (0.117) (0.103) (0.166) (0.337)

AIR21 × MC -0.016 0.008 0.006 -0.011 0.020 -0.004 0.014 0.027 0.034 0.064 (0.013) (0.015) (0.017) (0.021) (0.050) (0.011) (0.021) (0.019) (0.030) (0.065)

Observations 15,958 7,942 5,658 3,266 1,408 15,958 8,559 6,096 2,886 1,469 Firms 19 19 18 17 16 19 19 19 16 14 Markets 1,941 1,034 784 599 340 1,941 1286 1021 557 314 Firm / Market / Time FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Controls Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes R-squared 0.226 0.261 0.265 0.315 0.386 0.226 0.261 0.281 0.327 0.371 arbitrary threshold, and 0 otherwise. These placebo regressions are run on progres- sively narrower windows surrounding the arbitrary threshold. Using two alternative cutoffs of the forcing variable, 40% and 60%, I find no similar results as in the previous analyses, indicating that it is unlikely that my main results are driven by unknown effects that just happen to be present at the AIR-21 treatment threshold.

This is consistent with results from the non-parametric RD estimations in Fig- ure 1.4, where it is shown that there are no statistically significant discontinuities in price growth differentials at arbitrary thresholds. Together, these examinations bolster confidence on the validity of the effects of AIR-21 around the 50% cutoff.

1.4.5 Are Cash Rich Firms Constrained?

An important argument that has been made in the finance literature is that firms value and accumulate cash when they are financially constrained and lack access to external capital (see Almeida, Campello, and Weisbach(2004), Faulkender and

Wang(2006)). A potential issue for this paper, then, is that firms with more cash are actually those that are constrained, and therefore the reported effects of cash are capturing not the impact of financial flexibility but rather the lack of it. While the methodology of this paper attempts to circumvent the endogenous selection problem of firms’ cash holdings, it is nonetheless helpful to examine the effects of financial

flexibility implied by other measures to confirm that the cash effect is consistent with a flexibility story.

To address this issue, I show that firms that are less likely to be financially con- strained compared to their rivals based on alternative measures behave in the same

45 way as cash-rich firms do, whereas those more likely to be constrained do the oppo- site. Specifically, I first use ‘net cash’ as a more general measure of financial flexibility.

Since net cash is the amount of cash a firm has in excess of its debt obligations, firms with larger net cash balances should be less constrained. Next, I take cues from the

financial constraints literature and assume that high payout firms are likely to be unconstrained (see Fazzari, Hubbard, and Petersen(1988, 2000), Almeida, Campello, and Weisbach(2004) among others). Conversely, I take a reduction of dividend pay- ments in the previous year as a sign that a firm must have been constrained, given that cutting dividends is an expensive action that firms would avoid absent dire finan- cial straits (see DeAngelo and DeAngelo(1990), DeAngelo, DeAngelo, and Skinner

(1992)).22

In Table 1.9, I report results from triple difference regressions where cash is re- placed by the firm’s net cash position (i.e. cash minus total debt, scaled by total assets), payout ratio (i.e. dividends and repurchases scaled by total capital), and a dividend cut dummy (i.e. indicator for whether the firm had cut dividends in the pre- vious year). While Form 41 Financial Data, the primary source of firm-level financial data used in this paper, covers all airline companies that report to the Department of Transportation regardless of whether they are publicly listed or privately held, dividend payments or stock repurchases are not part of the reported items. For this reason, I use Compustat data to compute Payout and Dividend Cut which excludes privately held firms from the sample, dropping 6 out of 26 firms and truncating the

22Another source of flexibility for firms are lines of credit. Based on data from Capital IQ, however, I find little evidence that the use of credit lines by airline companies are commensurate with their cash holdings in magnitude or importance. Airlines on average hold roughly 15% of their assets in cash during my sample period, whereas their lines of credit amount to approximately 1% of assets with very little left undrawn. In untabulated analysis, I find that lines of credit have no impact on pricing and do not weaken the effects of cash when controlled for.

46 Table 1.9: Financial Flexibility vs. Constraints This table presents results from triple-difference regressions of 12Q price growth differentials (∆∆36Log(Market Fares)) on an AIR-21 indicator (AIR21), a set of alternative ex-ante relative- to-rival financial flexibility measures (Flexibility), ex-ante multimarket contact (MC), and their interactions. Financial flexibility measures include Net Cash ((cash - total debt) / total assets), Payout (from Compustat as (dividends (dvc + dvp)+ repurchases (prstkc))/at), Dividend Cut (an indicator variable equal to 1 if a firm had cut dividends in the previous year and 0 otherwise), and CashCompustat (from Compustat as che/at). Relative-to-rival flexibility and multimarket contact are constructed as described in Section 1.3, and are measured as of the previous year. Correlations between the flexibility measures are shown in Panel A. The regressions are run using the sample window 5% above and below the 50% AIR-21 cutoff. Results are reported in Panel B. Firm, market, and time fixed effects as well as control variables are included as in previous analysis. Standard errors adjusted for clustering at firm and market levels are reported in parentheses (*** p<0.01, ** p<0.05, * p<0.1).

Panel A. Correlations

Net Cash Payout Dividend Cut CashCompustat Net Cash 1 Payout 0.25 1 Dividend Cut -0.01 -0.11 1 CashCompustat 0.48 -0.07 -0.18 1 CashF orm41 0.51 -0.03 -0.19 0.85

Panel B. Regressions with Alternative Measures

Dependent Variable: ∆∆36Log(Market Fares) Compustat Sample

Flexibility (Constraint) Variables Net Cash Payout Dividend Cut CashCompustat

AIR21 × Flexibility -1.272*** -9.163** 0.407** -2.630*** (0.382) (3.385) (0.173) (0.713)

AIR21 × MC × Flexibility 0.237*** 1.650** -0.081** 0.528*** (0.074) (0.730) (0.034) (0.172)

MC -0.036 -0.012 -0.030 -0.023 (0.029) (0.033) (0.032) (0.036)

Flexibility 0.773*** 6.858*** -0.352*** 1.833** (0.228) (1.747) (0.062) (0.611)

MC × Flexibility -0.136** -1.358*** 0.061*** -0.353** (0.047) (0.433) (0.013) (0.118)

AIR21 -0.274** -0.267* -0.310** -0.278** (0.129) (0.130) (0.123) (0.128)

AIR21 × MC 0.046* 0.042 0.051** 0.045* (0.026) (0.024) (0.024) (0.025)

Observations 2,828 2,458 2,458 2,458 Firms 19 14 14 14 Markets 561 508 508 508 Firm / Market / Time FE Yes Yes Yes Yes Controls Yes Yes Yes Yes R-squared 0.343 0.355 0.357 0.356

47 sample period for a number of remaining firms. Net Cash, Payout, and Dividend

Cut are transformed into relative-to-rival values as described for Cash in previous sections.

Panel A demonstrates that relative-to-rival net cash, payout, and cash (computed from either Form 41 or Compustat) are all negatively correlated with the company’s status as a dividend cutter in comparison to its rivals. The correlation between cash and payout is negative, consistent with previous papers in the literature, but modest.

Cash and net cash are positively correlated since net cash is a linear combination of cash and debt. Assuring that using Form 41 or Compustat data does not make a notable difference, cash computed from Form 41 and Compustat are very highly correlated, and the correlations between cash and the alternative measures are similar regardless of which data is used to measure cash.

Panel B reports the regression coefficients. The triple difference regressions are run on the sample window where top-two airline passenger shares range from 45% to 55% (i.e. 5% above and below the 50% AIR-21 treatment threshold). The result from using net cash instead of cash is shown in the first column. Because leverage is already incorporated in computing net cash, I drop leverage and its interactions with AIR-21 and multimarket contact from the set of control variables in the first regression. The regression shows that financially flexible firms with high net cash respond to AIR-21 by pricing more aggressively. A Firm with one standard deviation

(i.e. 18.67 percentage point) greater ex-ante relative-to-rival net cash reduces its price growth rate over the next 36 months by 23.75 percentage points more from the last

36 months (i.e. 0.78 standard deviations lower price growth differential). The second and third columns present results from replacing cash with payout and the dividend

48 cut indicator. Consistent with the story that financial flexibility strengthens firms’ war chests, high payout firms respond to competition shocks by pricing more aggres- sively while firms that had cut dividends compete in a more accommodating manner.

Firms with one standard deviation (i.e. 1.35 percentage points) greater payout as a fraction of total capital relative to rivals exhibit 12.37 percentage points (i.e. 0.42 standard deviations) lower price growth differentials. In sharp contrast, a firm that had (not) cut dividends in the previous year while none (all) of its rivals had would raise (reduce) the rate of its price growth in the next 36 months by 40.7 percentage points more from the previous 36 months compared to a firm whose dividend cut status is the same as its rivals. The last column shows that using Compustat cash delivers similar results to the main analysis using cash from Form 41 Financial Data, despite the changes in firms and time periods composing the sample. A one standard deviation (i.e. 9.16 percentage point) increase in ex-ante relative-to-rival cash hold- ings as a fraction of assets leads to 24.09 percentage points lower price growth over the next 36 months compared to the previous 36 months (i.e. 0.81 standard deviations lower price growth differential). The effects of net cash, payout, dividend cut, and cash are canceled out by an increase in multimarket contact of 5.37%, 5.55%, 5.02%, and 4.98% respectively.

The evidence that supposedly constrained firms raise their price growth rates while financially flexible companies reduce them supports the notion that cash-rich

firms are able to price aggressively due to the financial flexibility provided by cash.

Another way to establish the flexibility story of cash is to demonstrate that it creates value. I do this next.

49 1.4.6 Cash and Market Performance Outcomes

To what extent is financial flexibility from cash valuable? Does the aggressive use of competitive firepower lead to economic benefit? A testable implication of the theory in this regard is that firms with financial resources at their disposal can afford to invest in longer term market shares at the expense of profits today (see Telser

(1966), Bolton and Scharfstein(1990)). I explore this by examining how ex-ante relative-to-rival cash positions and multimarket contact affect two metrics of market performance outcomes: market share gains and profitability growth.

To do this, I run the triple difference regressions using market share and prof- itability gains as alternative dependent variables. Market share gains are 36-month growth rates of firms’ market shares measured in terms of number of passengers and dollar revenues (i.e. average fare price × number of passengers). Profitability gains are 18-month (short-term) and subsequent 18-month (long-term) changes in firms’ profitability (i.e. return on assets (ROA)) attributable to each market. Firm-level profitability is attributed to each market by first obtaining market level operating profits as market revenue multiplied by the firms overall operating profit margin (i.e. operating profits/loss over operating revenues), and then dividing market level oper- ating profits/losses by the firm’s total assets. The regressions are then run on the sample window where top-two airline passenger shares range from 45% to 55% (i.e.

5% above and below the 50% AIR-21 treatment cutoff).

The results shown in Table 1.10 suggest that firms with more cash than their peers, which previous sections have shown to compete more aggressively, gain market shares by more (or lose them by less) over the course of 36 months. The coefficients

50 Table 1.10: Cash and Market Performance Outcomes This table presents results from triple-difference regressions of market performance outcomes (i.e. market share gains and profitability gains) on an AIR-21 coverage indicator (AIR21), ex-ante relative-to-rival cash holdings (Cash), ex-ante multimarket contact (MC), and their interactions. Market share gains are 12-quarter growth rates of the firm’s market shares measured in terms of number of passengers and dollar revenues (i.e. average fare price × number of passengers). Prof- itability gains are changes in the firm’s profitability (ROA) per market over the first 6-quarters and next 6-quarters. Profitability per market is computed by first obtaining market level operating profits as market revenue multiplied by the firm’s overall operating profit margin (i.e. operating profits/loss over operating revenues), and then dividing market level operating profits/losses by the firm’s total assets. The triple difference regressions are run using the sample window 5% above and below the 50% AIR-21 treatment cutoff. Firm, market, and time fixed effects as well as con- trol variables are included in all specifications as in previous analysis. Standard errors adjusted for clustering at firm and market levels are reported in parentheses (*** p<0.01, ** p<0.05, * p<0.1).

Dependent Variables Market Share Gain (12Q) Profitability (ROA) Gain Passengers Revenue First 6Q Next 6Q

AIR21 × Cash 2.190* 1.931* -0.255 0.537*** (1.153) (1.031) (0.261) (0.159)

AIR21 × MC × Cash -0.408* -0.390* 0.042 -0.102*** (0.224) (0.202) (0.044) (0.030)

MC -0.010 -0.018 -0.004 0.001 (0.065) (0.070) (0.007) (0.003)

Cash 0.274 0.729 0.369 -0.335** (1.135) (1.165) (0.269) (0.131)

MC × Cash -0.096 -0.189 -0.063 0.071** (0.208) (0.214) (0.045) (0.027)

AIR21 -0.466 -0.473 0.010 -0.021 (0.370) (0.383) (0.025) (0.023)

AIR21 × MC 0.077 0.081 -0.002 0.004 (0.061) (0.062) (0.005) (0.004)

Observations 2,828 2,828 2,781 2,781 Firms 19 19 19 19 Markets 561 561 553 553 Firm / Market / Time FE Yes Yes Yes Yes Controls Yes Yes Yes Yes R-squared 0.354 0.364 0.524 0.646

51 on the interaction term between AIR21 and Cash imply that a one standard devi- ation (i.e. 8.93 percentage point) increase in relative-to-rival cash leads to 19.56% and 17.24% greater market share growth based on number of passengers and dollar revenues, respectively. Moreover, while these firms appear to suffer mildly in terms of profitability in the shorter term (1-18 months), their profitability improves sig- nificantly by the time they gain market shares due to the strategic advantages of cash (18-36 months). Coefficients imply that in the nearer term after AIR-21 treat- ment, profitability drops by 2.28 basis points with a one standard deviation increase in relative-to-rival cash (though not statistically significant), but increases substan- tially and significantly by 4.8 basis points in the longer term, indicating that cash rich firms gain value in the long haul. Note that the magnitude of the profitability gain per market is non-trivial given that firms on average serve around 750 markets.

This provides further support for interpreting the cash effect on pricing as a financial

flexibility effect, and therefore complements the findings of Fresard(2010) who, while inconclusive about the underlying mechanism, also documents that cash rich firms gain more market shares after competition shocks.

Overall, the findings thus far are consistent with the hypothesis that financially

flexible firms are better equipped to engage in aggressive pricing strategies, and that the possibility of rival retaliation deters the competitive use of such flexibility. One question remains: If cash benefits firms under increased competition, do firms accu- mulate cash when competition unexpectedly intensifies?

52 1.4.7 The Effects of Competition on Corporate Cash Hold- ings

While the main focus of this paper is to establish a causal link from cash hold- ings to firm product pricing behavior, it is a more challenging task under my setting to claim causality from competition to cash holdings. The reason is that competi- tion shocks arrive at the market level while cash positions are a decision variable only at the firm level. There are 2,786 markets and 26 firms, so the market level is granular whereas the firm level is much less so. Furthermore, while it is plausible that firms may alter their pricing strategies depending on how much more financial slack they happen to have than their rivals, it is harder to argue that they would dynamically adjust their cash positions relative to their competitors in each market they serve. Therefore, it is not statistically feasible to test the impact of market level AIR-21 shocks on firm level cash holdings, and even meaningless to test the effects of market level competition shocks on firm-market level relative-to-rival cash holdings. Notwithstanding these challenges, a suggestive answer is given in Table 1.1 of Section 1.3 where cash holdings are shown to increase following the legislation of

AIR-21. A better visualization of the effect of market-level competition shocks im- posed by AIR-21 on corporate cash holdings is provided by Figure 1.7.

Here, I set the first date at which the origin airport of a market is covered by

AIR-21 as Quarter 0, and compute the passenger-weighted average cash-to-asset ra- tio of companies operating in each market for 12 quarters before and after the initial coverage date. Then, the cross-market passenger-weighted average of these within- market average cash holdings is plotted separately for the full sample, a restricted sample where the top-two airline passenger share determining coverage at Quarter

53 0.18

0.17

0.16

0.15

0.14

0.13 Cash / Assets 0.12

0.11

0.1

0.09 -12 -11 -10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6 7 8 9 10 11 12 Quarters

Within 10% of cutoff Within 15% of cutoff Full sample

Figure 1.7: Cash Holdings Before and After AIR-21 Coverage

This figure illustrates the impact of market-level AIR-21 competition shocks on airline company cash holdings. For each market, I set the first date at which the origin airport of the market is covered by AIR-21 as Quarter 0. I then compute the passenger-weighted average cash holdings (as a fraction of assets) of companies operating in each market for 12 quarters before and after the initial coverage date (from Quarter -12 to 12). I then plot the passenger-weighted average of these within-market average cash holdings across all markets for each relative quarter. Plots are shown separately for the full sample, a restricted sample where the top-two airline passenger share determining coverage at Quarter 0 lies within 15% of the 50% AIR-21 treatment cutoff (35% to 65%), and a restricted sample where the top-two airline passenger share determining coverage at Quarter 0 lies within 10% of the cutoff (40% to 60%).

0 lies within 15% of the 50% AIR-21 treatment cutoff (i.e. 35% to 65%), and a restricted sample where the top-two airline passenger share lies within 10% of the cutoff (i.e. 40% to 60%). Because markets will have varying initial coverage dates, this approach alleviates the concern that shifts in cash holdings are simply due to some unobserved time effect. The plots from narrow window samples also strengthen

54 the argument that the depicted changes in cash holdings are due to AIR-21 shocks.

In Figure 1.7, I observe a sharp increase in passenger-weighted average cash hold-

ings after initial AIR-21 coverage, consistent with the idea that firms rationally re-

spond to competition shocks by accumulating their financial war chests, which in turn

enable them to price aggressively in the face of such shocks in the future.23 While

the jump in cash holdings is substantial (i.e. nearly a 50% increase from 10% to

15% of total assets within a year) and therefore strongly supportive of the impact of

AIR-21 competition shocks, it should be cautioned that this is by no means proof

of a causal impact since there is no counterfactual trend of cash holdings that the

plotted trend can be compared against. The set of firms operating in markets never

covered by AIR-21 are in fact almost identical to the set of firms operating in markets

covered by AIR-21 at some point. Nonetheless, Figure 1.7 suggests that competition

dynamics maybe a potential determinant of corporate cash holdings.24

In the last section, I make some concluding remarks.

1.5 Conclusion

Can cash be a source of financial flexibility that enables firms to compete aggres- sively in product markets? Yes, especially when there is less concern of retaliation

23Note that although firm cash holdings increase on average, there can be heterogeneity in the degree of such responses depending on the potential likelihood of rival retaliation faced by each firm (captured by the level of multimarket contact). 24There is a very large literature on the determinants of corporate cash holdings. Among several papers, Opler, Pinkowitz, Stulz, and Williamson(1999) and Bates, Kahle, and Stulz(2009) suggest cash flow volatility to be a key determinant of corporate cash holdings. Haushalter, Klasa, and Maxwell(2007) come closer to the spirit of my argument that competition dynamics are also a likely determinant of cash holdings. They find that the extent to which firms have interdependent growth prospects with rivals (i.e. face higher predation risk) is positively associated with cash holdings and the use of hedging derivatives. Hoberg, Phillips, and Prabhala(2014) also show that product market threats significantly influence cash retention decisions.

55 from rivals (i.e. when there is little multimarket contact). In this paper, I investigate

one mechanism through which corporate cash holdings affect product market out-

comes: firm pricing strategy. I test whether firms with larger cash reserves relative

to their rivals price more aggressively as ‘deep pocket’ theories predict (see Telser

(1966), Bolton and Scharfstein(1990)), and whether greater concern of potential ri-

val retaliation (i.e. higher multimarket contact) attenuates this competitive effect of

cash as implied by theories from industrial organization (see Bernheim and Whinston

(1990), Evans and Kessides(1994)).

I turn to the airline industry where markets and rivalry are cleanly defined. To

circumvent concerns that cash and multimarket contact are endogenously linked to

firm pricing strategy, I exploit an industry-wide regulation, the Aviation Investment

and Reform Act for the 21st Century (AIR-21), to identify exogenous market-level competition shocks and test the effects of ex-ante relative-to-rival cash holdings and multimarket contact on ex-post changes in firm pricing policy. Under AIR-21, large commercial airports whose two largest airlines enplane more than 50% of the air- port’s total passengers are required to submit and implement competition enhance- ment plans under the supervision of the Federal Aviation Administration (FAA). The features of this regulation facilitate a regression discontinuity (RD) design that miti- gates endogeneity concerns and ensure that firms are not able to predict or self-select into AIR-21 treatment.

In triple difference regressions on progressively narrower windows surrounding the

50% treatment threshold, I find that firms with greater cash holdings relative to rivals respond to market-level competition shocks by pricing more aggressively, particularly

56 when multimarket contact is sufficiently low (i.e. when there is less concern of retal- iation from rivals). A one standard deviation, or 8.55 percentage point, increase in relative-to-rival cash as a fraction of assets one year prior to an AIR-21 competition shock in a market leads to 15.22 percentage points lower price growth over the next 36 months compared to the previous 36 months (i.e. 0.47 standard deviations lower price growth differential). Multimarket contact has a sizable impact on this strategic effect of cash: a 5.3% increase would completely overturn the cash effect. Consistent with the competitive impact of AIR-21 mainly being driven by airport gate reallocations toward low-cost carriers (LCCs), I find that LCCs respond to AIR-21 competition shocks by pricing aggressively irrespective of their cash holdings, while legacy airlines respond aggressively conditional on holding more cash. In both cases, the responses are dampened by higher multimarket contact. Also, the main results hold only for

firms that had high ex-ante market share (i.e. firms for which AIR-21 indeed serves as a competition shock), but are non-existent for firms that had low market share to begin with (i.e. firms for which AIR-21 rather serves as an accommodative event).

To alleviate concerns that firms might predict AIR-21 coverage in a given mar- ket and build-up cash reserves relative to rivals in that market in advance, I use relative-to-rival cash measured 3 and 4 years prior to treatment (one year prior in baseline specifications) and show that results are robust. Placebo tests using alterna- tive threshold levels of top-two airline passenger shares as AIR-21 treatment cutoffs confirm that the main results are unlikely due to other confounding effects that hap- pen to coincide with competition shocks induced by AIR-21. I further cement the role of cash as a source of financial flexibility by showing that firms with high net cash or high payout ratios compete aggressively in response to AIR-21 as do cash-rich

57 firms, and that supposedly constrained firms that had cut dividends in the previous year behave in an opposite accommodating manner. I also demonstrate that such

flexibility is valuable by showing that firms holding more cash than rivals gain more market shares and experience long-term profitability growth.

How corporate cash holdings are used in the firm’s day-to-day operations are of central interest both in the academic literature and among policy makers. This paper contributes to the growing body of research exploring the interplay between finan- cial flexibility and product market competition by documenting how cash affects firm pricing strategies, thereby highlighting the mechanism through which cash impacts product market outcomes. This study also opens the door to further exploration of how competition dynamics shape corporate cash decisions. I look forward to future research broadening our understanding in this dimension.

58 Chapter 2: Hedge Fund Activism and Internal Capital Markets

2.1 Introduction

Since the early 2000s, hedge funds have joined the fray of shareholder activism, transforming both the means and ends of exacting influence as owners of corporations.

With the decline of activity in the market for corporate control in the 1980s, share- holder initiated proxy proposals and interventions, particularly those carried out by large institutional investors, have become an important aspect of shareholder moni- toring. Among such institutional shareholders, hedge funds, despite having received much public spotlight and suspicion for their confrontational tactics, have stood out as successful at imposing lasting change on the value and operating performance of target companies (see Brav, Jiang, Partnoy, and Thomas(2008), Bebchuk, Brav, and

Jiang(2015), and Brav, Jiang, and Kim(2015)). How do they bring about such change? An important yet relatively unanswered question is how and to what extent hedge funds influence firms’ inner operations in the long-run.25 This paper aims to

narrow this gap by exploring how interventions from activist hedge funds impact the

25Only recently has research started to shed light on this aspect. See, notably, Brav, Jiang, and Kim(2015) for evidence on the positive impact of hedge fund activism on the production efficiency of target firms.

59 efficiency of the firm’s internal capital allocation decisions.

A widely debated proposition in finance is that diversified firms are valued less than undiversified firms, and that this discount is due to the inefficient investment allocation across divisions (see Lang and Stulz(1994), Berger and Ofek(1995), and

Shin and Stulz(1998)). If this is the case, hedge fund activism may affect the internal capital markets of target companies for several reasons. A direct reason is that hedge funds care about the firm’s investment policies as part of their goal to improve the operational efficiency of the firm, and therefore intervene so that internal capital is directed toward projects with the greatest investment opportunities. Brav, Jiang,

Partnoy, and Thomas(2008) report that among 1,059 activist events between 2001 and 2006, improving operational efficiency is stated as the main goal in 131 of the cases. For example, Brav, Jiang, and Kim(2015) show that firms targeted by activist hedge funds invest in technologies that increase the productivity of their plants. Fur- thermore, hedge funds are concerned with alleviating agency problems and divisional power struggles within the firm, which are potentially at the roots of inefficient capital allocation across divisions (see Rajan, Servaes, and Zingales(2000) and Scharfstein and Stein(2000)). Hedge funds may indirectly affect the efficiency of internal capi- tal markets by addressing such political and social frictions within the organization, particularly when they do not have the expertise to intervene in the firm’s technolo- gies, operations, and assets. Consistent with this notion, Klein and Zur(2009) show that hedge funds focus on addressing agency problems associated with cash flows, in which sense hedge funds have similar governance oriented objectives as other activist shareholders (see Hartzell and Starks(2003) and Ertimur, Ferri, and Muslu(2011)).

Whichever the underlying rationale may be, it is reasonable to expect that a more

60 efficient internal capital market may arise as a result of activist campaigns by hedge funds, if they indeed intend to increase the value of the firm as they purport.

Alternative views in the conglomerates literature challenge the assertion that in- ternal capital markets are inefficient, by arguing that the empirically observed diver- sification discount and internal capital allocations are consistent with optimal firm behavior (see Maksimovic and Phillips(2002) and Gomes and Livdan(2004)); that the diversification discount is related to the reduction of risk and uncertainty about the firm’s prospects rather than operational inefficiencies (see Mansi and Reeb(2002), and Hund, Monk, and Tice(2010)); or that the conclusions about internal capital market efficiency and the value of diversified firms are products of selection biases or measurement errors (see Whited(2001), Campa and Kedia(2002), Graham, Lem- mon, and Wolf(2002), Villalonga(2004), and Cust´odio(2014)). If there really is no inefficiency in the way resources are allocated within conglomerates, interventions by activist hedge funds should have little impact on how diversified firms use their internal capital markets. At the least, significant shifts in resource allocations follow- ing activist campaigns would imply that hedge funds take the view that the internal capital markets of their target companies are inefficient.

A good example illustrating the intent and consequences of hedge fund activism is the case of Pershing Square Capital Management, an activist hedge fund led by Bill

Ackman, and Canadian Pacific Railway, a major railroad company in which the hedge fund acquired a 14% stake at the end of 2011. In early 2012, Mr. Ackman launched an activist campaign, pushing for changes in the company. His complaints were focused on the lack of operating efficiency, low asset utilization, substantial underinvestment, failed acquisitions, and sluggish market share growth. Mr. Ackman argued that the

61 incumbent CEO and the CEO-friendly board were inept at addressing such problems, and pushed for change of management with the full support of the shareholders of the company. After the new CEO, industry legend Hunter Harrison, and board members chosen by Mr. Ackman were brought in, significant operational improvements took place, including substantially more investment allocations to areas with high growth opportunities such as intermodal services and the firm’s rail network covering North

Dakota’s Bakken shale region. From September 2011 to December 2014, Canadian

Pacific’s stock price rose from around $49 to $220.

Overall, theory and evidence provide tension for the hypothesis that the internal capital markets of firms targeted by activist hedge funds may become more active and efficient. To investigate this hypothesis, I implement a difference-in-differences frame- work to analyze the differential impact of hedge fund activism on segment investment with respect to the cash flows and investment opportunities of business segments. As in previous studies of hedge fund activism, I rely on Schedule 13D filings submitted by hedge funds to the SEC to identify targeted companies. Many of these papers examine short-term stock price movements surrounding the 13D filing date, or study the actions of hedge funds and their impact on the performance of target companies over a short period of time mostly before they exit their investments (see Brav, Jiang,

Partnoy, and Thomas(2008), Klein and Zur(2009), Brav, Jiang, and Kim(2015), and Gantchev, Gredil, and Jotikasthira(2015)). In contrast, I abstract from the in- vestment horizon of hedge funds and analyze the long-term impact that they have on the internal capital markets of targeted companies.

Consistent with the hypothesis that activist hedge funds push firms to facilitate and optimize internal capital markets, I find that investments made by the segments

62 of targeted firms become significantly more sensitive to cash flows generated in other parts of the company after it first becomes an investment target in a Schedule 13D

filing, and that most of this increase in cross-subsidization is driven by reallocating the firm’s cash flows toward segments with high Tobin’s Q. The sensitivity of a typical segment’s investment to cash flow generated by other segments of the firm increases by 0.046, much of which is contained among segments with the highest Tobin’s Q in the firm: It increases by 0.056 more for the highest Q segments than for segments that have lower Q. I interpret the results as evidence that hedge funds push firms to rectify their internal capital markets such that cash flows generated in various segments of the firm are redirected toward the segment with the greatest investment opportunities.

To mitigate the concern that segments with the highest Tobin’s Q do not correctly identify those with the greatest investment opportunities due to measurement errors in Tobin’s Q (see Whited(2001)), I show that the results are robust to categorizing segments with respect to their Tobin’s Q in several ways. Separately estimating the effect of hedge fund activism on cross-subsidization using dummy variables for (1) highest Q segments, (2) above median Q segments, (3) above average Q segments, and (4) lowest Q segments within firms all corroborate the conclusion that hedge funds influence firms to redirect cash flows to segments with the greatest invest- ment opportunities. The results also hold for an industry-size matched sample and a propensity score matched sample, making it less likely that inferences about the effects of hedge fund activism are confounded by systematic differences between tar- geted and non-targeted firms. Finally, I alleviate the concern that estimated changes

63 in cross-subsidization may in fact be a manifestation of changes in unobservable cor- relations across the investment opportunities of segments (see Chevalier(2004)), by showing that there is little evidence that targeted firms refocus their businesses by reducing the number of their business segments or industries they serve.

This paper contributes to the literature on active monitoring by outside share- holders, and the nascent field of hedge fund activism in particular. Earlier studies cast doubt on the effectiveness of activist campaigns in the 1980-1990s carried out by institutional shareholders such as mutual funds and pension funds. Karpoff, Malat- esta, and Walkling(1996), Smith(1996), Wahal(1996), and Gillan and Starks(2000), for instance, suggest that the impact of interventions by such activist institutions on

firm value and operating performance are modest or non-existent. This has recently changed with the advent of activist hedge funds. Brav, Jiang, Partnoy, and Thomas

(2008), Bebchuk, Brav, and Jiang(2015), and Brav, Jiang, and Kim(2015) collec- tively show that interventions by hedge funds lead to increased operating performance,

CEO turnover, and production efficiency in target firms, which are accompanied by large abnormal returns that are not reversed in the near future. Moreover, Gantchev,

Gredil, and Jotikasthira(2015) find that hedge fund activism has spillover effects on non-targeted firms as well. This paper sheds more light on the real effects of their interventions.

The results of the paper also provide new insight into the workings of internal cap- ital markets, namely that outside shareholders can influence the efficiency at which they operate. Much focus in the internal capital markets literature has been placed on their general lack of efficiency (see Shin and Stulz(1998)), and ample theory

64 and evidence have been put forth explaining the role of agency conflicts and intra-

firm politics therein (see Rajan, Servaes, and Zingales(2000), Scharfstein and Stein

(2000), Gertner, Powers, Scharfstein(2002), Ozbas and Scharfstein(2010), Duchin and Sosyura(2013), and Glaser, Lopez-de-Silanes, and Sautner(2013)). My findings suggest that the fates of internal capital markets can be altered by force of actively monitoring shareholders. While more research can be done on this front, the results of this paper are consistent with activist hedge funds removing social and political barriers within the firm that prevent capital from being allocated efficiently.

In Section 2.2, I discuss the empirical strategy of the paper. Key results are presented in Section 2.3. Finally, I conclude in Section 2.4.

2.2 Empirical Strategy

2.2.1 Data

To investigate the effects of activist interventions from hedge funds on the ef-

ficiency of internal capital markets, I examine whether cross-divisional investment subsidies within the firm become more aligned with division-level investment oppor- tunities after a firm is targeted by an activist hedge fund.

Hedge fund activism is identified using Schedule 13D filings, or “beneficial owner- ship reports”, submitted to the SEC. Section 13(d) of the 1934 Securities Exchange

Act stipulates that investors who (1) own more than 5% of a voting class of a com- pany’s equity securities and (2) intend to influence control of the issuer must disclose the amount and intent of ownership within 10 days of acquiring such a stake. The investor can file a shorter 13G filing in lieu of a 13D in the absence of the intent to control, which implies that a Schedule 13D filing meaningfully indicates an active

65 intervention to follow. 13D filers are narrowed down to hedge fund managers based on the identity descriptions of the reporting entities. I then take a firm to be tar- geted by hedge fund activism, with a dummy variable set equal to 1 (HFA= 1) and 0 otherwise, if a Schedule 13D had initially been filed with the firm as the investment target at least a year earlier.

The sample of my study covers the Compustat universe of firms that report seg- ment level information over the period 1996 to 2012.26 Segment level financial data are obtained from Compustat business segment files, and firm level information are supplemented from the Compustat fundamentals annual database. Segment level in- vestment and cash flows are both normalized by total assets of the firm, since a dollar of cash flow should have the same impact on segment investment after controlling for investment opportunities, regardless of where the cash flow originates within the

firm.27 Detailed variable descriptions and screening procedures to address various reporting biases are elaborated in the data appendix.

Table 2.1 presents a snapshot of the data. There are 875 (2,413) firms (segments) of which 95 (280) are targeted by activist hedge funds.28 Targeted firms have an average asset size of $ 2,065 million, while non-targeted firms have $ 5,412 million.

This sharp difference in firm size is consistent with the notion that hedge funds are

26I start the sample period in 1996 to ensure a minimum gap of two years with 1994, the year of the first available 13D filing date, so that biases arising from unobserved activist events before 1994 are balanced with a long enough sample period. 27To the extent that firms can tap into their internal capital markets when they are credit- constrained, the cash flow of a segment should affect the investment of the firm only through its impact on firm cash flow (see Shin and Stulz(1998)). 28Due to the segment and industry-year fixed effects included in the baseline specification, I require that there are more than one observation for each segment, and more than one segment within each industry-year.

66 Table 2.1: Sample Statistics This table presents summary statistics for the final sample over the period 1996 to 2012. The number of firms and segments are shown for the full sample, the subsample of firms targeted by hedge fund activism, non-targeted firms, and non-targeted control samples matched on industry (exact match on 2-digit SIC code) and firm size or propensity scores (closest match). Propensity scores are predicted values from a logistic regression of a hedge fund activism dummy (HFA) on lagged firm level Tobin’s Q, cash flow, asset size, cash holdings, and long-term debt. Average firm characteristics such as asset size, cash holdings over assets, long-term debt over assets, return on assets, cash flow, Tobin’s Q, sales growth, capital expenditure over assets, within-firm Herfindahl-Hirschman indices of segment level sales, and number of segments are presented. t-statistics on the difference of means between the targeted vs non-targeted sample and targeted vs matched control samples are shown as well. The number of activist hedge funds in the sample and their Herfindahl-Hirschman indices in terms of target asset size and number of targets are shown at the bottom.

Full Non- Size P-Score Sample Treated Treated (t-stat) Matched (t-stat) Matched (t-stat) Firms 875 95 780 149 254 Segments 2,413 280 2,133 437 753

Firm Characteristics Size 4,908 2,065 5,412 (-4.23) 1,525 (1.60) 2,168 (-0.32)

Cash 0.07 0.07 0.07 (-0.72) 0.08 (-1.04) 0.08 (-1.96)

Debt 0.24 0.26 0.23 (3.01) 0.26 (-0.07) 0.22 (3.33)

ROA 0.01 0.01 0.01 (0.84) 0.02 (-0.70) 0.02 (-0.58)

Cash Flow 0.12 0.12 0.12 (0.57) 0.13 (-1.32) 0.12 (0.09)

Tobin’s Q 1.47 1.40 1.49 (-1.99) 1.43 (-0.40) 1.41 (-0.13)

Sales Growth 0.23 0.13 0.24 (-0.78) 0.15 (-0.42) 0.14 (-0.34)

Capital Expenditure 0.06 0.06 0.07 (-2.07) 0.06 (-0.46) 0.06 (-0.37)

HHI of Segment Sales 0.44 0.43 0.44 (-0.85) 0.44 (-0.44) 0.45 (-1.00)

Number of Segments 2.87 2.78 2.89 (-2.05) 2.69 (1.36) 2.79 (-0.09)

Funds 115 HHI (Target Size) 0.10 HHI (No. of Targets) 0.06

less likely to target bigger firms because the large amount of capital it takes to ac- quire a 5% stake in a large company could introduce significant portfolio risk. Firms that become targets tend to hold smaller cash balances, be more indebted, be more

67 profitable, have greater cash flow, have lower Tobin’s Q, have lower sales growth, invest less, and be more diverse in terms of how dispersed their segment sales are.29

Targeted firms also tend to have fewer segments, consistent with them being smaller

firms. Furthermore, there are 115 distinct hedge funds that invest in these diversified

firms, and the Herfindahl-Hirschman index of funds based on either target size or tar- get numbers is very small (0.10 and 0.06, respectively), indicating that it is unlikely the case that a small number of major funds might drive the results of the analysis.

2.2.2 Endogeneity

While observable firm characteristics are explicitly controlled for in my analy- sis, the significant differences between the firm characteristics of targeted and non- targeted firms raise the concern that the distribution of these characteristics may be different between the two groups. One may then suspect that they may also be inherently different in unobservable ways. If that is the case, the observed effects of hedge fund activism on internal capital markets may potentially include confounding influences from those unobservable dimensions that are correlated with hedge fund activism. The primary reason this endogeneity is a concern is that it then becomes unclear whether it is the action of hedge funds that causes internal capital markets to become more active and efficient, or if fund managers simply choose to invest in

firms where such changes are imminent even without any activist intervention.

Prior research on the effects of hedge fund activism largely refutes the possibility that the latter might be the case on average. Notably, hedge fund activist campaigns are associated with changes in target companies that are unlikely to take place in

29Lang and Stulz(1994) and Comment and Jarrell(1995), for instance, use the Herfindahl index computed from segment sales as a measure of firm diversification.

68 the absence of pressure from activists, such as a sharp increase in the firm’s CEO turnover rate, from less than 6% to over 12% over the 2001 to 2006 period (see Brav,

Jiang, Partnoy, and Thomas(2008)). Activist campaigns also entail significant costs for the activist investor: Campaigns that end in proxy fights are estimated to cost on average near $11 million (see Gantchev(2013)). Furthermore, studies that compare announcement returns of 13D filings with 13G filings document higher returns for activist stake disclosures than for passive ones (see Clifford(2008) and Klein and Zur

(2009)). These findings cannot be explained if the average activist hedge fund engages in passive stock picking rather than active engagement. In essence, the evidence to date suggests it is unlikely that the endogeneity problem will seriously confound the results of this study.

Econometrically, it is a difficult if not infeasible task to rule out the stock picking story by estimating the average treatment effect of hedge fund activism in a random- ized experiment, because hedge funds are likely to choose their battles in firms where there is a clear problem they want to solve and where they can readily effect change to their liking. In the absence of an instrument for hedge fund activism or an exoge- nous shock thereof, a partial remedy is to use a matched sample where targeted and non-targeted firms are reasonably similar so that treatment by hedge fund activism is the only source of observable variation across the two groups. The hope is that once the two groups are made to be similarly distributed in observable ways, other sys- tematic variations will be minimized, except for whether a firm is targeted by hedge fund activism. For this purpose, I construct two matched samples that employ only non-targeted firms that are similar to target firms as the control group. Each year,

69 targeted firms are matched to non-targeted companies in the same two-digit SIC in- dustry that are closest according to some metric. In the first matching method, firms are matched by asset size, where the difference (computed as |ATtargeted−ATcontrol| ) is (ATtargeted+ATcontrol)/2 required to be smaller than 50%. In the second method, firms are matched by their propensity scores, which are computed as the predicted values from a pooled logis- tic regression of a hedge fund activism dummy (HFA) on a set of lagged firm level variables (i.e. Tobin’s Q, cash flow, asset size, cash holdings, and long-term debt).

These industry-size or propensity score matched non-targeted firms are then used as the control group.

The last four columns of Table 2.1 show the characteristics of these matched firms.

There are 149 (254) and 437 (753) industry-size (propensity score) matched control

firms and segments, respectively. What is notable is that across the various firm characteristics, there is no longer a significant difference between matched control

firms and targeted firms. For example, the average asset size of non-targeted firms drops from $ 5,412 million to $ 1,525 million after the industry-size matching, and the difference with targeted firms, which is significant with a sizable t-statistic of

-4.23 before matching, becomes insignificant (t-statistic=1.60). Characteristics other than asset size are also well matched as a result of the industry-size matching scheme.

The difference in long-term debt (t-statistic goes from 3.01 to -0.07), Tobin’s Q (-

1.99 to -0.40), capital expenditure (-2.07 to -0.46), and number of segments (-2.05 to 1.36) all become insignificant after matching on industry and size. This suggests that the industry-size match does a good job at making targeted and non-targeted

firms comparable. The propensity score matching scheme also does a reasonable job

70 at minimizing the differences between targeted and non-targeted firms for most char- acteristics, with the exception of long-term debt. I conduct my analysis using the full sample as well as the industry-size and propensity score matched samples, and report results for each case.

2.2.3 Methodology

I employ these samples to test the hypothesis that hedge fund activism makes the sensitivity of a division’s investment to cash flow generated elsewhere in the firm more responsive to whether the division has greater opportunities than other seg- ments of the firm. To this end, I implement a difference-in-differences regression of segment investments (INVS) on an indicator variable for hedge fund activism treat- ment (HFA), a set of explanatory variables meant to capture the extent and efficiency of investment cross-subsidies across divisions, and the interaction terms between those variables and the HFA dummy variable. These explanatory variables include (1) the segment’s own cash flow, (2) cash flows from all the other segments within the firm, and (3) their interactions with an indicator variable for whether the segment has the greatest investment opportunities among all segments of the firm.

An important hurdle in this approach is the measurement of segment level in- vestment opportunities. While Tobin’s Q is conventionally used as a measure of investment opportunities for firms, it is not explicitly calculable for segments be- cause segment market values are unavailable. As widely done in the conglomerates literature to circumvent this challenge, I use the median Tobin’s Q of “pure-play” single-segment firms in the same industry as the segment, defined at the two-digit

SIC level, as a proxy for the segment’s Tobin’s Q (see Shin and Stulz(1998), Rajan,

71 Servaes, and Zingales(2000), and Ozbas and Scharfstein(2010) for notable papers

that employ this methodology).

This method by itself does not resolve the potential bias due to the measurement

error in Tobin’s Q. As discussed by Whited(2001) and many others, observable mea- sures of Tobin’s Q can diverge substantially from the actual unobservable measure

of investment opportunities implied by standard intertemporal models of investment:

marginal Q. In the context of the difference-in-differences framework used in this

study, this can be a problem if the measurement error in Tobin’s Q distorts the or-

dering of investment opportunities of segments within firms, or is somehow correlated

with whether a firm is targeted by hedge fund activism. While it is difficult to imag-

ine why measurement errors would correlate with hedge fund activism, it is plausible

that such errors could result in falsely categorizing a segment as having the highest

investment opportunities when in fact it does not.

One way to address this issue is to define whether a segment has high or low

Tobin’s Q compared to other segments of the firm in different ways, and check if

the results from those alternative Q categorizations convey a consistent message. I

alternatively identify segments that have Tobin’s Q greater than the firm’s median

and average and show similar results with identifying those with the highest Tobin’s

Q. Furthermore, identifying segments with the lowest Tobin’s Q yields sharply op-

posite results. These mitigate the concern that a segment might be falsely identified

as having the greatest investment opportunities in the firm because of measurement

errors. Also, I use lagged segment sales growth as an additional control measure of

segment investment opportunities that is based on data from the segment.

A related issue that often makes it difficult to estimate the extent of internal

72 capital market transfers is that apparent cross-subsidization may in truth be a mani-

festation of unobserved correlations of investment opportunities across segments (see

Chevalier(2004)). Using Tobin’s Q and sales growth to explicitly control for seg- ment investment opportunities partially alleviates this problem, only to the extent that they accurately measure investment opportunities. The difference-in-differences framework overcomes this limitation and further remedies the problem, insofar as the relatedness between segments do not change substantially after firms are targeted by hedge fund activism. However, it is a legitimate concern that hedge funds may indeed influence how close segments are to each other by making the firm more fo- cused, rendering the difference-in-differences specification invalid for the purpose of testing internal capital market activity. After all, it is well known from previous research that activist shareholders often push for the divestiture of underperforming assets (see Bethel, Liebeskind, and Opler(1998), Brav, Jiang, Partnoy, and Thomas

(2008), Klein and Zur(2009), Bebchuk, Brav, and Jiang(2015), and Brav, Jiang, and

Kim(2015)). In the next section, however, I show that activist interventions do not appear to be associated with significant changes in the number of business segments operated by firms or industries served by them. It appears that while hedge funds do push firms to divest unprofitable assets, they seldom go so far in eventuality as to shut down or sell off entire industry segments.

73 The difference-in-differences specification can be written as:

INVSi,j,t = α + β · HFAi,t

+ δ1 · CFi,−j,t + δ2 · CFi,j,t

+ ϕ1 · CFi,−j,t × HFAi,t + ϕ2 · CFi,j,t × HFAi,t

+ γ1 · (CFi,−j,t · HighQi,j,t−1) + γ2 · (CFi,j,t · HighQi,j,t−1)

+ λ1 · (CFi,−j,t · HighQi,j,t−1) × HFAi,t

+ λ2 · (CFi,j,t · HighQi,j,t−1) × HFAi,t

0 0 + µ · Xi,j,t−1 + τ · Yi,t−1 + ai,j + bt + i,j,t

where INVSi,j,t denotes gross investment of the jth segment of firm i during year

t, scaled by firm i’s total assets as of year t − 1. HFAi,t indicates whether it has

been at least 1 year as of year t since firm i had first become a target of hedge fund

activism. CFi,j,t, or CF, denotes the cash flow of segment j of firm i during year t, scaled by t − 1 total assets as of the firm. CFi,−j,t, or Other CF, denotes the cash

flows of all segments of firm i other than segment j. HighQi,j,t−1 indicates whether the segment has the highest Tobin’s Q among all segments of the firm. Xi,j,t−1 is a vector of lagged segment level control variables such as segment Tobin’s Q and sales growth, and Yi,t−1 includes firm level controls such as asset size, profitability, cash holdings, and leverage. Segment and industry-year fixed effects are controlled for as well.

The baseline specification captures the activity and efficiency of internal capital markets, and more importantly, the impact of hedge fund activism on both of these aspects. For example, the coefficient δ1 quantifies the increase in segment level in- vestment with respect to an incremental increase in cash flows generated elsewhere in the firm, hence the level of activity of internal capital markets. γ1 measures how

74 greater the investment cross-subsidy is when the segment has the greatest investment

opportunities in the firm. This coefficient assesses how efficient the firm’s internal

capital market is by estimating how much more capital is allocated toward the great-

est opportunities.

Most critically, ϕ1 and λ1 capture the impact of hedge fund activism on the activ-

ity and efficiency of internal capital markets, respectively. ϕ1 evaluates the effect of

hedge fund activism on the sensitivity of segment level investment to cash flow from

other segments in the firm. λ1 allows for a different coefficient on the interaction of

other segments’ cash flow and HFA for segments with the highest Q in the firm. The

idea is that if activist hedge funds have a real effect on the capital allocation between

divisions of target companies, one would expect either of the coefficients ϕ1 or λ1 to be economically large and statistically significant.

In the next section, I explain the key results from implementing this methodology on the full sample as well as on the industry-size matched sample and the propensity score matched sample.

2.3 Key Results

Table 2.2 shows the full sample results from the difference-in-differences frame- work described in the previous section. In the first and second columns, I exclude the treatment indicator for hedge fund activism (HFA) and analyze the effects of a segment’s cash flow and cash flow from other segments on the segment’s investment in an OLS framework, controlling for the segment’s investment opportunities measured by Tobin’s Q and sales growth. Consistent with Shin and Stulz(1998), I find that increases in both the segment’s own cash flow and other segments’ cash flow lead

75 Table 2.2: Hedge Fund Activism and Internal Capital Markets This table presents results from difference-in-differences regressions of segment investments (INVS) on a treatment indicator variable for whether the firm is a target of hedge fund activism (HFA), a set of explanatory variables: other segments’ cash flow (Other CF ); own segment cash flow (CF ); their interactions with a dummy variable indicating whether the segment has the highest Tobin’s Q among all segments of the firm (High Q), and the interaction terms between these explanatory variables and HFA. All specifications include segment and industry-year fixed effects, where industry is defined as the segment’s 4-digit SIC code. Segment Q and sales growth, and firm asset size (log of assets), profitability (ROA), cash holdings and long-term debt (fraction of assets) are included as control variables. Standard errors adjusted for clustering at the firm-year level are reported in parentheses. (*** p<0.01, ** p<0.05, * p<0.1)

Dependent Variable: INVS

Other CF × High Q × HFA 0.056** (0.025)

CF × High Q × HFA -0.007 (0.046)

Other CF × HFA 0.046** 0.028 (0.023) (0.024)

CF × HFA -0.044 -0.038 (0.032) (0.035)

HFA 0.004 0.003 (0.004) (0.004)

Other CF × High Q -0.001 -0.005 (0.009) (0.009)

CF × High Q 0.003 0.005 (0.018) (0.019)

Other CF 0.027*** 0.022** 0.017* 0.023** 0.019* (0.010) (0.010) (0.010) (0.010) (0.010)

CF 0.085*** 0.080*** 0.084*** 0.078*** 0.082*** (0.013) (0.013) (0.014) (0.015) (0.016)

No. Obs. 7,893 7,893 7,893 7,893 7,893 Segment Controls Yes Yes Yes Yes Yes Firm Controls No Yes Yes Yes Yes Segment / Industry-Year FE Yes Yes Yes Yes Yes Adj R2 0.699 0.703 0.704 0.702 0.704

76 to significantly more segment investment, but that the effect of the segment’s own cash flow is much larger in magnitude. The coefficient on the segment’s own cash flow

(0.080) is 3.6 times larger than that on other segments’ cash flow (0.022). The central question is whether this relative inactivity of cash flow transfers across segments, or internal capital market failure, is resolved by the intervention of activist hedge funds, and if so, whether the resolution is attained by increasing the efficiency of internal capital markets.

The difference-in-differences result reported in the third column of Table 2.2 indi- cates that hedge fund activism has an effect of facilitating internal capital markets.

The coefficient on the interaction term between other segments’ cash flow (Other CF ) and the hedge fund activism dummy variable (HFA) implies a positive and significant impact of hedge fund activism on the sensitivity of segment investment to other seg- ments’ cash flow: INVS-to-OtherCF sensitivity increases on average by 0.046 (with a

firm-year clustered standard error of 0.023). On the other hand, hedge fund activism does not appear to affect the sensitivity of the segment’s investment to its own cash

flow. Taking into account the fact that a segment’s investment is much more sensitive to its own cash flow than to other segments’ cash flow to begin with, the evidence cor- roborates the idea that hedge funds come in to remove forces that prevent resources from flowing from one segment to another. Moreover, the interaction term between other segments’ cash flow and hedge fund activism partially subsumes the effect of other segments’ cash flow alone, which means that internal capital markets are even less active without such activist interventions.

Why are internal capital markets inactive in the first place, and how do hedge funds help facilitate them? The canonical answer to the first question provided by

77 the literature is that internal capital markets fail because they do not actively direct

internal resources of the firm to their best use (see Shin and Stulz(1998), Rajan, Ser-

vaes, and Zingales(2000), Ozbas and Scharfstein(2010), Duchin and Sosyura(2013),

and Glaser, Lopez-de-Silanes, and Sautner(2013)). The fourth column of Table 2.2

shows results consistent with this understanding. Here, the segment’s own cash flow

and other segments’ cash flow are interacted with an indicator variable for whether

the segment’s Tobin’s Q is the highest in the firm (High Q). The coefficient indicates

there is no evidence that even the segment with the greatest investment opportunities

receives more internal resources than any other segment does. The crux is, do hedge

funds make firms rectify this?

To examine how firms targeted by activist hedge funds activate internal capital

markets, the last column of Table 2.2 interacts the HFA dummy with the segment’s

own cash flow, other segments’ cash flow, and those for the highest Q segments sep-

arately. Confirming that hedge funds induce firms to redirect cash flows toward the

greatest investment opportunities within the firm, the coefficient for the interaction

term between Other CF, High Q, and HFA is significantly positive. Following hedge

fund activism, the sensitivity of segment investment to other segments’ cash flow for

the highest Q segments increases on average by 0.056 (with a firm-year clustered standard error of 0.025) more than for segments that do not have the highest Tobin’s

Q. More importantly, it subsumes the coefficient on the interaction term between

Other CF and HFA, providing evidence that hedge funds facilitate the internal cap-

ital markets of firms predominantly by making them more efficient. They push the

firm’s internal capital market to redirect cash flows generated in various segments

toward the segment with the highest Tobin’s Q.

78 As mentioned earlier in the paper, a potential problem with Tobin’s Q-based measures of investment opportunities is that they may contain measurement error.

Consequently, the High Q indicator variable might falsely assign a value of 1 to a segment that is in fact not the one with the highest marginal Q. If Tobin’s Q hap- pens to diverge greatly from marginal Q, this will be a problem. I address this issue by using alternative definitions of High Q: The segment’s Tobin’s Q is (1) above the median Tobin’s Q of the firm’s segments; (2) above the average Tobin’s Q of

firm segments; (3) lowest among all segments of the firm. The first two alternative definitions allow more room for error in estimated Tobin’s Q, because there are less likely to be errors in judging whether a segment has greater investment opportuni- ties than most segments of the firm, compared to judging whether the segment has greater opportunities than all the other segments. The third alternative definition, that the segment has the lowest, not highest, Tobin’s Q, corroborates the baseline results by providing evidence going in the opposite direction. The combined results alleviate concerns about measurement errors, since it is unlikely that high Tobin’s Q coincides with low marginal Q and at the same time low Tobin’s Q corresponds to high marginal Q, on average.

Table 2.3 shows the results from this analysis. In the first and second columns, I define High Q to equal 1 if the segment has Tobin’s Q higher than the median and average of the firm’s segments, respectively, and 0 otherwise. The results are similar to that shown in the last column of Table 2.2. In both cases, the coefficient on the interaction term between other segments’ cash flow, High Q, and HFA is larger in magnitude: 0.068 (significant at the 5% level) and 0.065 (significant at the 1% level)

79 Table 2.3: Internal Capital Market Efficiency: Alternative High Q Defini- tions

This table presents results from diff-in-diff regressions of segment investments (INVS) on a hedge fund activism dummy (HFA), a set of explanatory variables: other segments’ cash flow (Other CF ); own segment cash flow (CF ); their interactions with a High Q dummy, and the interaction terms between these explanatory variables and HFA. Results are presented for three alternative definitions of High Q: whether the segment’s Tobin’s Q is (1) above the median segment’s Tobin’s Q; (2) above average; (3) lowest among all segments of the firm. All specifications include segment and industry- year fixed effects. Segment Q and sales growth, and firm asset size, profitability, cash holdings, and long-term debt are included as control variables. Standard errors are adjusted for clustering at the firm-year level. (*** p<0.01, ** p<0.05, * p<0.1)

Alternative High Q Definitions Dependent Variable: INVS Above Median Above Average Lowest Q

Other CF × High Q × HFA 0.068** 0.065*** -0.078*** (0.027) (0.025) (0.026)

CF × High Q × HFA 0.011 0.008 0.001 (0.045) (0.044) (0.044)

Other CF × HFA 0.017 0.024 0.072*** (0.025) (0.024) (0.027)

CF × HFA -0.049 -0.045 -0.046 (0.037) (0.036) (0.038)

HFA 0.003 0.003 0.004 (0.004) (0.004) (0.004)

Other CF × High Q -0.011 -0.006 0.006 (0.009) (0.008) (0.009)

CF × High Q -0.013 -0.002 0.018 (0.019) (0.018) (0.020)

Other CF 0.023** 0.020* 0.014 (0.011) (0.010) (0.010)

CF 0.091*** 0.085*** 0.074*** (0.017) (0.016) (0.017)

No. Obs. 7,893 7,893 7,893 Control Variables Yes Yes Yes Segment / Industry-Year FE Yes Yes Yes Adj R2 0.704 0.704 0.704

80 for the above median and above average High Q definitions, respectively. As in Ta-

ble 2.2, hedge fund activism impacts the sensitivity of segment investment to other

segments’ cash flow only for High Q segments, but not for below median and below

average Q segments. In the third column, I assign segments with the lowest Tobin’s

Q in the firm with a value of 1 and assign 0 for all other firms. Not surprisingly, hedge

fund activism has a much lower impact on the sensitivity of segment investment to

other segments’ cash flow for the lowest Q segments, compared to the impact it has

on other segments of the firm with better prospects. The coefficient on the interaction

term between other segments’ cash flow and HFA is 0.072 which is highly significant at the 1% level, and the coefficient differs by -0.078 (also significant at 1%) for the lowest Q segments. In essence, the intervention of hedge funds ensure that resources are directed toward the most promising divisions and not to places where internal capital may be wasted.

Another complication discussed earlier is that targeted and non-targeted firms may be systematically different from each other. This confounds inference about the causal effect of hedge funds, because hedge funds may be passively picking companies whose internal capital markets are likely to become more efficient in the future rather than actively effecting change. In the econometric sense, it is difficult to disentangle this problem. However, because hedge fund activism is usually associated with high campaign costs to the activist investor and changes in the company that are likely involuntary (e.g. higher CEO turnover), the endogeneity concern is a benign issue in the context of this study. Notwithstanding, a feasible and partial econometric remedy is to construct a matched sample so that targeted and non-targeted companies are

81 Table 2.4: Matched Sample Difference-in-Differences Regressions This table presents matched sample results from diff-in-diff regressions of segment investments (INVS) on a hedge fund activism dummy (HFA), a set of explanatory variables: other segments’ cash flow (Other CF ); own segment cash flow (CF ); their interactions with a High Q dummy, and the interaction terms between these explanatory variables and HFA. Each year, targeted firms are matched with non-targeted firms in the same two-digit SIC industry with closest asset size (Panel A) or propensity scores (Panel B). Propensity scores are predicted values from a logistic regression of a hedge fund activism dummy (HFA) on lagged firm level Tobin’s Q, cash flow, asset size, cash holdings, and long-term debt. Results are presented using alternative definitions of High Q: whether the segment’s Tobin’s Q is (1) the highest; (2) above median; (3) above average; (4) lowest in the firm. All specifications include segment and industry-year fixed effects, as well as firm and segment level control variables. Standard errors are adjusted for clustering at the firm-year level. (*** p<0.01, ** p<0.05, * p<0.1)

Panel A. Industry-Size Matching Panel B. Propensity Score Matching Alternative High Q Definitions Alternative High Q Definitions Above Above Above Above Dependent Variable: INVS Highest Q Median Average Lowest Q Highest Q Median Average Lowest Q

Other CF × High Q × HFA 0.114** 0.109* 0.118** -0.111** 0.092* 0.096* 0.106** -0.098* (0.051) (0.057) (0.052) (0.056) (0.049) (0.054) (0.051) (0.059)

82 CF × High Q × HFA -0.036 -0.014 -0.051 -0.015 -0.064 -0.048 -0.072 0.014 (0.076) (0.092) (0.088) (0.087) (0.075) (0.083) (0.083) (0.088)

Other CF × HFA 0.091* 0.044 0.040 0.045 0.148** 0.080* 0.044 0.032 0.038 0.135** (0.047) (0.046) (0.052) (0.047) (0.060) (0.042) (0.044) (0.046) (0.044) (0.059)

CF × HFA -0.011 -0.002 -0.003 0.010 -0.007 -0.064 -0.027 -0.034 -0.025 -0.063 (0.057) (0.060) (0.074) (0.067) (0.066) (0.052) (0.055) (0.065) (0.061) (0.064)

HFA -0.010 -0.006 -0.009 -0.008 -0.009 -0.001 -0.001 -0.002 -0.000 -0.002 (0.008) (0.008) (0.008) (0.008) (0.008) (0.007) (0.007) (0.007) (0.007) (0.007)

Other CF × High Q -0.013 -0.018 -0.007 0.030 -0.010 -0.004 -0.004 0.019 (0.031) (0.037) (0.032) (0.028) (0.031) (0.037) (0.032) (0.035)

CF × High Q 0.125** 0.065 0.124* -0.028 0.052 0.025 0.055 0.023 (0.054) (0.065) (0.067) (0.053) (0.063) (0.070) (0.069) (0.059)

Other CF -0.051 -0.051 -0.050 -0.054 -0.069 -0.036 -0.040 -0.042 -0.044 -0.052 (0.045) (0.047) (0.051) (0.047) (0.044) (0.048) (0.052) (0.053) (0.052) (0.050)

CF 0.029 -0.019 -0.010 -0.022 0.048 0.032 -0.000 0.008 -0.001 0.022 (0.066) (0.070) (0.078) (0.071) (0.074) (0.044) (0.043) (0.048) (0.044) (0.063)

No. Obs. 879 879 879 879 879 1,017 1,017 1,017 1,017 1,017 Control Variables Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Segment / Industry-Year FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Adj R2 0.749 0.755 0.750 0.755 0.749 0.744 0.745 0.745 0.746 0.744 made to be similar at least across observable dimensions with the hope that system- atic differences between the two groups will also be minimized. As elaborated in the previous section, I construct an industry-size matched sample as well as a propensity score matched sample. Each year, targeted companies are matched with non-targeted companies in the same two-digit SIC industry with closest asset size or propensity scores, which are used as the control sample. As shown in Table 2.1, the matching procedures successfully minimize differences between the targeted and non-targeted groups across a variety of observable variables. I then implement the difference-in- differences framework on each of these matched samples.

Table 2.4 shows results from this matched sample analysis. The bottom-line is that the results from the full sample hold up well in both of the matched samples.

One notable difference is that the segment’s own cash flow and other segments’ cash

flow alone no longer account for the segment’s investment policy. The coefficients on

CF and Other CF range from -0.022 to 0.048 and from -0.069 to -0.036, respectively, none of which are statistically significant. This is consistent with the story that hedge funds tend to target firms where there are more impediments to active resource re- allocation, and since the control samples are made to have similar characteristics to the targeted sample, the overall level of internal capital market activity is low for

firms in the matched samples. The lack of segment investment sensitivity to the

firm’s cash flow can also be explained by the fact that targeted firms tend to invest less than non-targeted firms, potentially underinvesting as often accused by activist hedge funds, and therefore the matched samples which are constructed to be similar to targeted firms exhibit lower investment to cash flow sensitivity.

The first columns in Panel A (industry-size match) and Panel B (propensity score

83 match) of Table 2.4 show that hedge fund activism has a sizable impact on how sen-

sitive a segment’s investment is to cash flows generated elsewhere within the firm.

While the statistical significance is somewhat weaker than in the full sample results,

potentially due to the substantially smaller sample size, the economic magnitude is

large: a 0.091 (0.080) increase in the segment investment sensitivity to other seg-

ments’ cash flow for the industry-size (propensity score) matched sample. As shown

in the remaining four columns in each panel of Table 2.4, the efficiency implications

are profound in both of the matched samples. Again, the segment investment to

other segments’ cash flow sensitivity gains are statistically significant and econom-

ically large for segments that have the highest Tobin’s Q in the firm, but not for

other segments. For example, the coefficient on the interaction term between other

segments’ cash flow, High Q, and HFA is a striking 0.114 (with a firm-year clustered

standard error of 0.051) in the industry-size matched sample, and the interaction

term between other segments’ cash flow and HFA is insignificant. This result holds

up when using alternative definitions of High Q. The coefficient on OtherCF × HFA is greater by 0.109 (0.118) for firms with above median (average) Tobin’s Q compared to firms that have lower Tobin’s Q. In the last column of each panel, I confirm that segments with the lowest Tobin’s Q in their firms receive substantially less cash flows from other segments compared to segments that have higher Tobin’s Q, after being targeted by hedge fund activism. For instance, the sensitivity of segment investment to other segments’ cash flow in the industry-size matched sample increases dramati- cally by 0.148 after being targeted, but lowest Q segments are excluded from enjoying that cross-subsidization. The difference in the sensitivity increase between lowest Q

84 Table 2.5: Hedge Fund Activism and Firm diversification This table presents results from logit and probit regressions of dummy variables for whether a firm reduced or increased the number of reported segments or number of segment industries in a given year, on a treatment indicator variable for whether the firm had been a target of hedge fund activism (HFA). Industry is defined at the 2-digit industry level. Firm Q, cash flow, sales growth, asset size, profitability, cash holdings, and long-term debt are included as control variables. Year dummies are controlled for in all specifications. Marginal effects are presented as results. Results are reported for the full sample and industry-size matched sample separately. Standard errors are adjusted for clustering at the firm level. (*** p<0.01, ** p<0.05, * p<0.1)

Logistic Regressions Probit Regressions ∆Seg. ∆Seg. ∆SIC ∆SIC ∆Seg. ∆Seg. ∆SIC ∆SIC < 0 > 0 < 0 > 0 < 0 > 0 < 0 > 0 Panel A. Full Sample

HFA -0.009 -0.014* -0.012 -0.011* -0.009 -0.015** -0.012 -0.010* (0.008) (0.008) (0.008) (0.006) (0.008) (0.007) (0.007) (0.006)

No. Obs. 3,878 3,670 3,130 3,700 3,878 3,670 3,130 3,700 Pseudo R2 0.055 0.107 0.040 0.068 0.054 0.108 0.039 0.067 Prob > χ2 0.004 0.000 0.060 0.000 0.004 0.000 0.090 0.000

Panel B. Industry-Size Matched Sample

HFA 0.018 0.001 0.004 0.003 0.020* 0.001 0.006 0.004 (0.012) (0.020) (0.016) (0.015) (0.012) (0.018) (0.015) (0.013)

No. Obs. 460 368 303 379 460 368 303 379 Pseudo R2 0.272 0.170 0.234 0.081 0.272 0.179 0.234 0.087 Prob > χ2 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000

Panel C. Propensity Score Matched Sample

HFA 0.010 0.001 0.001 -0.004 0.011 0.000 0.001 -0.004 (0.014) (0.017) (0.013) (0.010) (0.012) (0.015) (0.011) (0.009)

No. Obs. 550 475 428 554 550 475 428 554 Pseudo R2 0.264 0.141 0.243 0.081 0.262 0.145 0.236 0.084 Prob > χ2 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000

Control Variables Yes Yes Yes Yes Yes Yes Yes Yes Year Dummy Yes Yes Yes Yes Yes Yes Yes Yes

segments and other segments is both economically large (-0.111) and statistically sig- nificant (at 5%).

Finally, I provide ancillary evidence that it is not the case that firms targeted by

85 hedge funds become more focused on average. This suggests that while hedge funds may push for the divestiture of unprofitable assets, they rarely close entire industry segments in actuality.30 This is an important point for this study, because if firms become more focused due to hedge fund activism, it may well be the case that the apparent increase in cross-subsidization following hedge fund intervention is in fact a product of increased unobserved correlation between segment investment opportu- nities.31 In Table 2.5, I report results from logistic and probit regressions analyzing whether firms change the number of segments or segment industries after being tar- geted by hedge fund activism. I analyze four possibilities. The dependent variable is a dummy variable set to 1 if the firm (1) reduced the number of its business segments;

(2) increased the number of its business segments; (3) reduced the number of seg- ment industries; (4) increased the number of segment industries during the year, and

0 otherwise. Industry is defined at the 2-digit SIC level. The independent variable is an indicator for whether the firm has been the target of activist hedge funds for at least a year (HFA). Firm Tobin’s Q, cash flow, sales growth, asset size, profitability, cash holdings, long-term debt, and year dummies are included as control variables.

Results are shown separately for the full sample (Panel A), industry-size matched sample (Panel B), and propensity score matched sample (Panel C). The coefficients are reported as marginal effects and standard errors are adjusted for clustering at the

firm level.

Contrary to the concern that hedge funds may push the firm to become more

30See Bethel, Liebeskind, and Opler(1998), Brav, Jiang, Partnoy, and Thomas(2008), Klein and Zur(2009), Bebchuk, Brav, and Jiang(2015), and Brav, Jiang, and Kim(2015) for previous research that suggest that hedge funds aim to divest underperforming parts of target companies. 31The concern that cross-subsidization measured from segment cash flows may in fact be due to correlated investment opportunities across segments was raised by Chevalier(2004).

86 focused and therefore confound inferences about their impact on cross-subsidization, the marginal effects of hedge fund activism on the probability of reducing the num- ber of business segments or the number of segment industries are both economically and statistically negligible across the board. Out of the eight specifications test- ing whether the number of segments or industries decrease following hedge fund ac- tivism, the only indication that hedge funds make firms more focused is given by an industry-size matched sample probit regression of whether firms reduce the number of segments. The marginal effect is a 2% increase in probability which is significant at the 10% level. On the other hand, there is slightly more evidence that targeted

firms become less likely to further diversify their businesses. Four specifications using the full sample yield significant results, though the magnitude of the marginal effects are modest. For example, logit (probit) regressions show that targeted firms are 1.4%

(1.5%) and 1.1% (1.0%) less likely to increase the number of their business segments and the number of industries in which they operate, respectively. Overall, Table 2.5 shows that hedge fund activism does not appear to be associated with large changes in the configuration of business segments of target companies, mitigating the possi- bility that the estimated effects of hedge fund activism on internal capital markets are due to changes in unobservable cross-segment correlations.

In the next section, I summarize the paper and provide concluding remarks.

2.4 Conclusion

This paper sheds new light on the impact of hedge fund activism on the inner workings of firms, namely that the intervention of hedge funds push firms to facili- tate their internal capital markets and improve the efficiency at which they allocate

87 resources across business segments. This is an important finding given the large liter-

ature documenting the limits of internal capital markets and the social barriers that

prevent them from working properly. It also provides insight into the widely debated

role hedge funds play in the companies they launch activist campaigns against.

I find that following initial Schedule 13D filings, investments made by the seg-

ments of targeted firms become substantially more sensitive to cash flows generated

in other parts of the company, and that most of the increase in cross-subsidization

comes from the reallocation of firm cash flows toward segments with high Tobin’s Q.

These findings are robust to categorizing segments with respect to their Tobin’s Q in a variety of ways, mitigating the concern that measurement errors in Tobin’s Q might drive the results. Allowing for separate estimation of the effect of hedge fund activism on cross-subsidization for (1) highest Q segments, (2) above median Q segments, (3) above average Q segments, and (4) lowest Q segments all deliver a consistent message that hedge funds redirect the firm’s cash flow to segments with the greatest invest- ment opportunities. The results also hold in an industry-size matched sample as well as a propensity score matched sample, rendering it less likely that systematic differ- ences between targeted and non-targeted firms confound inferences about the effects of hedge fund activism. Finally, there is no evidence that targeted firms refocus their businesses by reducing the number of their business segments or industries they serve, alleviating the concern that apparent changes in cross-subsidization may in fact be a symptom of changes in unobservable correlations across segments.

It should be noted that the findings of this paper are silent about the precise rationale behind the facilitation of internal capital markets that hedge funds achieve.

88 It could be the case that hedge funds are concerned about productivity, and there- fore directly impact the effectiveness of the firm’s use of capital and assets. On the other hand, hedge funds may correct inefficiencies indirectly by preventing political frictions between CEOs and divisional managers from distorting resource allocations.

This can particularly be true when hedge funds, as outside investors, are not experts about the firm’s detailed operations. The truth is likely somewhere in between.

Activist campaigns carried out by hedge funds are often at the center of public contention, where debates frequently deviate from the facts. With corporate control and shareholder wealth at stake, it is important for academics to understand the real implications of hedge fund activism regarding the operations of firms. While our knowledge has advanced in recent years, there is still much room for research to understand how activist hedge funds affect the utilization of labor and capital within the firm, and how they influence the social connections and internal power struggles of divisional managers. Such endeavors may also help understand how hedge funds differ from other types of shareholders who actively seek to influence corporate policies. I look forward to fruitful investigations in the future.

89 Chapter 3: (Priced) Frictions

3.1 Introduction

Microstructure frictions cause the observed price of a stock to deviate from the fundamental “true” price. Examples of such frictions include bid-ask bounce caused by intermediation, stale prices caused by non-synchronous trading, orders originating with uninformed momentum traders pushing price away from fundamental value, or temporary price pressure from large orders. The effects of these frictions are non- trivial. They can lead to significant differences between average observed returns and true expected returns, which academic researchers and investment professionals often assume to be synonymous in their analysis. They are also costly for investors who may demand compensation for bearing the frictions. Nonetheless, little has been done to accurately and comprehensively measure the impact of these frictions, and often- times naive methods are used in research as well as investment practice to mitigate the impact of stocks with potentially large frictions.

In this paper, we propose a parsimonious measure to characterize the severity of microstructure frictions at the individual stock level that is not only easy to construct but also theoretically motivated. It is simple to construct because all that is required to compute it is daily stock returns, and thus can be easily applied in settings where

90 other microstructure-based proxies are more difficult to obtain. Our measure is moti- vated by the key insight in Blume and Stambaugh(1983) that the noise in observed prices induced by microstructure frictions causes (due to Jensen’s inequality) the av- erage observed return to be upward biased. More importantly, they show that the magnitude of the bias is equal to the variance of the frictions at the beginning of the return holding period. We exploit this property to construct a “friction-free” measure of the average daily return of a firm by dividing its average two-day observed total return by the average lagged one-day observed return. Since the holding periods for the two-day return and the lagged one-day return begin at the same point in time, both returns are affected by the same amount of frictions, which are cancelled out in the division, thus leaving a consistent friction-free estimate of the firm’s average daily return. The same observation has been exploited by Fisher and Weaver(1992),

Fisher, Weaver, and Webb(2010), and Asparouhova, Bessembinder, and Kalcheva

(2010, 2013) to correct for microstructure friction-induced biases in average portfolio returns.32 By contrast, we use the difference between a firm’s simple average daily observed return and its friction-free average daily return to measure the severity of the microstructure frictions faced by the firm, which we call FRIC, and then use this firm-specific measure to study the impact of microstructure frictions on the cross section of stock returns.

It is important to note that our friction measure captures microstructure effects due to any transient deviations of prices from fundamental true values (e.g. bid-ask bounce, non-synchronous trading, orders originating with uninformed traders, and

32Fisher and Weaver(1992) and Fisher, Weaver, and Webb(2010) correct for biases in index returns and Asparouhova, Bessembinder, and Kalcheva(2010, 2013) correct for biases in equally- weighted portfolios and cross-sectional regressions.

91 temporary price pressure from large orders). Hence, our measure provides a very simple yet comprehensive proxy of various microstructure frictions at the individual stock level, while at the same time being theoretically motivated.

We find that firms with the largest microstructure frictions are low priced, small, and volatile. Not surprisingly, the friction measure is also correlated with liquidity and other microstructure-based variables commonly used in the literature. Specifically, we find that FRIC is highly correlated with the Amihud(2002) illiquidity measure, bid-ask spread estimator (CS) from Corwin and Schultz(2012), Zero Return Pro- portion (ZP) measure from Lesmond, Ogden, and Trzcinka(1999), dollar volume, share turnover, bid-ask spread used in Amihud and Mendelson(1986) and Brennan,

Chordia, and Subrahmanyam(1998), and updated Gibbs estimates (C) used in Has- brouck(2009). In addition, we show that the market-wide average of FRIC displays considerable variation over time, with clearly visible spikes during the recessions of

1974 to 1975 and 1990 to 1991, the tech bubble of the late 1990s, and the financial crisis of 2007-2009.

Using the FRIC measure, we examine the pricing of microstructure frictions in the cross section of returns. Do investors require compensation for investing in firms with large microstructure frictions? The answer is yes. We find that, on a value-weighted basis, firms in the highest FRIC decile command a return premium as large as 0.82% per month in raw returns and 0.76% per month after adjusting for exposures to com- mon risk factors. The friction premium can reach as high as 1.27% per month in raw returns and 1.32% per month in risk-adjusted returns when stocks are weighted equally.

92 The friction premium is robust to alternative variable constructions, return ad- justments using different factor models, and subperiod and subsample analysis. In particular, we perform a number of robustness checks related to the way we construct our friction measure. Our primary measure is based on estimating a friction-free av- erage daily return using the average two-day observed return and the average lagged one-day observed return, which may lead to an underestimation of the bias in observed returns for stocks that trade infrequently. To investigate the extent of this potential downward bias and how it affects our results, we estimate alternative friction-free av- erage daily returns using longer windows (three-day vs. lagged two-day and four-day vs. lagged three-day) and find that our results are robust. We also vary the horizon of past daily returns used to estimate the friction measure. Our baseline results rely on one year of past daily returns, but we redo the analysis using six months, three years, and five years of past daily returns and find similar results.

As a robustness check to portfolio sorts and time-series factor regressions, we use Fama and MacBeth(1973) cross-sectional regressions to examine the return pre- dictability of our friction measure. In addition to a large number of standard controls such as size, book-to-market, momentum, profitability, and asset growth, these re- gressions also include many traditional liquidity price impact and cost measures that are correlated with our friction measure. The results from the Fama-MacBeth regres- sions show that the return explanatory power of our measure remains robust after controlling for those standard return predictors as well as traditional liquidity proxies.

In fact, the regression results suggest that our friction measure is a stronger cross- sectional return predictor than most of the traditional liquidity variables.

We also explore the interaction between the friction premium and known return

93 predictors. We find that the friction premium is much larger among small, value,

volatile, high accrual, and low asset growth firms with low past returns and prof-

itability. In addition, the friction premium is more pronounced among illiquid firms

as defined by traditional liquidity proxies. This result is consistent with the notion

that the impact of microstructure frictions is most severe among illiquid firms.

We further demonstrate the difference between our measure and conventional

proxies of frictions by examining how they affect the inferences on well-known asset

pricing anomalies. We do so in the context of the momentum (Jegadeesh and Titman

(1993, 2001)) and idiosyncratic volatility (Ang, Hodrick, Xing, and Zhang(2006))

anomalies. Empirical studies of these anomalies typically screen out firms with low

prices (Jegadeesh and Titman(2001) and Bali and Cakici(2008)) to mitigate the

impact of microstructure frictions. The choice of price filters is motivated by ease of

use combined with a lack of sufficiently long time-series data on more direct proxies of

microstructure frictions. Now, our friction measure is very easy to compute as it only

requires daily stock returns, and can go back as far as the 1920s when CRSP daily

data begin. To illustrate the differential impact of our friction measure, we show

how the momentum and idiosyncratic volatility (IVOL) anomalies are affected by

screening out the same number of firm/month observations with high FRIC instead

of low prices (Price < $1 or Price < $5). We find that the 10-1 decile return spreads

associated with momentum and IVOL are uniformly smaller in magnitude and have

lower t-statistics under FRIC filters than under corresponding price filters. Further- more, the Hou, Xue, and Zhang(2015) q-factor model performs better in explaining

the momentum and IVOL anomalies after screening out high FRIC firms instead of

low price firms. We also conduct quintile double-sorts first on price/FRIC and then

94 on momentum/IVOL to investigate how our friction measure compares to price level in identifying where the momentum and IVOL anomalies are concentrated and more generally in capturing microstructure frictions. We find that there is no momentum or IVOL anomaly in the lowest (highest) price (FRIC) quintile. More importantly, both anomalies tend to increase in magnitude with decreasing FRIC (and in the case of momentum monotonically so), reaching the highest level in the lowest FRIC quin- tile. In contrast, they exhibit an inverted U-shaped relation with price level, being most pronounced for the middle price quintiles. Therefore, we conclude that FRIC is a very different measure of microstructure frictions from price level, and that sorting on FRIC produces a much cleaner pattern in asset pricing anomalies than sorting on price level.

Finally, as an out-of-sample application, we repeat the analysis using stock mar- ket data from the UK. As discussed earlier, an important advantage of our friction measure is that it requires only daily stock returns to construct, thus avoids utiliz- ing variables that may not be available under certain circumstances. For instance, liquidity proxies such as the Amihud(2002) illiquidity measure, dollar volume, and share turnover require data on trading volume, the quality or availability of which may be suspect in settings such as the emerging markets or over-the-counter markets

(see, e.g., Bekaert, Harvey, and Lundblad(2007)), or may not be comparable across exchanges such as between NYSE and Nasdaq. Our analysis shows that the friction measure is also significantly priced for UK firms, and more importantly, it is the only variable among the list of microstructure liquidity proxies that uncovers a consistent return premium in the cross section of UK stock returns.

The remainder of the paper is organized as follows. Section 3.2 uses a simple

95 model to illustrate and motivate our friction measure. Section 3.3 describes the data,

and explains how we construct our friction measure using daily stock returns. Sec-

tion 3.4 presents summary statistics on friction-sorted portfolios, and the correlations

between our friction measure and other microstructure liquidity proxies. Section 3.5

and 3.6 study the impact of our friction measure on the cross section of stock returns

using portfolio sorts and Fama-MacBeth regressions, respectively. Section 3.7 studies

the interactions between the friction premium and other firm characteristics that are

known to predict returns. Section 3.8 uses our friction to examine the impact of

microstructure frictions on the momentum and idiosyncratic anomalies. Section 3.9

demonstrates the usefulness and robustness of our measure using stock market data

from the UK. Section 3.10 concludes.

3.2 Simple Model

3.2.1 Basic Setup

Our measure is based on the insight that the microstructure friction-induced devi-

ation of the observed stock price from the fundamental true price results in an upward

bias in the average returns of individual stocks. In this section, we present a simple

model to demonstrate this idea. The setup of our model largely follows that of Blume

and Stambaugh(1983). Specifically, the observed price of a stock is modeled as

Pbi,t = (1 + ϕi,t) Pi,t (3.1)

where Pbi,t is the observed price for firm i at time t which can deviate from the true

price, Pi,t, at which buy and sell orders would simultaneously cross absent transac- tional frictions. The noise term that captures this deviation, ϕi,t, is mean zero and independently distributed across t, and is also independent of Pi,τ for all τ. Note that

96 the noise term captures various microstructure frictions that cause observed prices to

deviate from fundamental true value such as bid-ask bounce, non-synchronous trad-

ing, orders originating with uninformed traders, and temporary price pressure from

large orders.

Assuming no dividends, the true (gross) return of security i for period t is simply defined as

Pi,t 1 + ri,t = (3.2) Pi,t−1 The observed return is, using (3.1), defined as

Pbi,t (1 + ϕi,t) Pi,t 1 + rbi,t = = (3.3) Pbi,t−1 (1 + ϕi,t−1) Pi,t−1 Combining (3.2) and (3.3) yields the following relationship between the observed

and true returns:

1 + ϕi,t 1 + rbi,t = [1 + ri,t] (3.4) 1 + ϕi,t−1 Taking expectations on both sides of (3.4) gives   1 + ϕi,t E [1 + rbi,t] = E E [1 + ri,t] (3.5) 1 + ϕi,t−1

Approximating the first term on the right hand side of (3.5) by a second order

Taylor series expansion and assuming that ϕi,t is serially uncorrelated, we have

 1 + ϕ  E i,t ≈ 1 + σ2 (3.6) ϕi,t−1 1 + ϕi,t−1

Substituting (3.6) into (3.5) yields

h i E [1 + r ] ≈ 1 + σ2 E [1 + r ] (3.7) bi,t ϕi,t−1 i,t

Ignoring the cross product term, σ2 E [r ], we can rewrite equation (3.7) as ϕi,t−1 i,t

E [1 + r ] ≈ E [1 + r ] + σ2 (3.8) bi,t i,t ϕi,t−1 97 Hence, the bias in the observed return induced by microstructure frictions is ap-

proximately equal to the variance of the noise term at the beginning of the holding

period, time t − 1. We exploit this property, that the bias is unrelated to the noise at the end of the holding period, in constructing our friction measure.

3.2.2 Methodology: Constructing “FRIC”

We first define a stock’s two-period expected return ending at time t, denoted

E [1 + ri,t−2,t], as

E [1 + ri,t−2,t] = E{[1 + ri,t−2,t−1] [1 + ri,t−1,t]} (3.9)

where the one-period expected return ending at time t, E [1 + ri,t], is now denoted as E [1 + ri,t−1,t]. We allow for the true return process to be serially correlated ri,t =

µi + ρi · ri,t−1 + i,t, with an autocorrelation coefficient of ρi and an idiosyncratic innovation term, i,t, and assume that this innovation in true returns is independent of the noise term capturing market frictions in observed prices, ϕi,t. With these assumptions, the covariance term in equation (3.9), Cov (ri,t−2,t−1, ri,t−1,t), is equal to

ρ σ2 / (1 − ρ2). Then, we can rearrange (3.9) as i i i

ρ σ2 i i E [1 + ri,t−2,t] E [1 + ri,t−1,t] = 2 + (3.10) (1 − ρi ) E [1 + ri,t−2,t−1] E [1 + ri,t−2,t−1]

Note that the observed two-period return can be written in terms of the true two-period return and noise terms as

Pbi,t (1 + ϕi,t) Pi,t 1 + ϕi,t 1 + rbi,t−2,t = = = [1 + ri,t−2,t] (3.11) Pbi,t−2 (1 + ϕi,t−2) Pi,t−2 1 + ϕi,t−2

Taking expectations of (3.11) and approximating by Taylor series yields,

h i E [1 + r ] ≈ 1 + σ2 E [1 + r ] (3.12) bi,t−2,t ϕi,t−2 i,t−2,t

98 which is analogous to steps (3.5) through (3.7). Similarly, taking expectations of the lagged observed one-period return ending at time t − 1, we have

h i E [1 + r ] ≈ 1 + σ2 E [1 + r ] (3.13) bi,t−2,t−1 ϕi,t−2 i,t−2,t−1

Dividing (3.12) by (3.13), then, results in the following equation, where the last step derives from equation (3.10):

h 2 i 1 + σ E [1 + ri,t−2,t] E [1 + rbi,t−2,t] ϕi,t−2 E [1 + ri,t−2,t] ≈ h i = E [1 + rbi,t−2,t−1] 1 + σ2 E [1 + r ] E [1 + ri,t−2,t−1] ϕi,t−2 i,t−2,t−1 (3.14) ρ σ2 i i = E [1 + ri,t−1,t] + 2 (1 − ρi ) E [1 + ri,t−2,t−1] Equation (3.14) shows that dividing the expected value of the two-period observed return by that of the lagged one-period observed return gives the true expected one- period return with an additional term due to the serial correlations of the true return process. Importantly, the noise in observed prices resulting from microstructure fric- tions is no longer a factor in equation (3.14). This is because both the expected two-period observed return and expected lagged one-period observed return share the same amount of noise, 1+σ2 . Therefore, dividing these two effectively cancels ϕi,t−2 out the common noise from microstructure frictions. In other words, equation (3.14) provides us with a “friction-free” estimate of the expected one-period true return.

Recall that equation (3.8) shows that the expected one-period observed return, denoted E [1 + rbi,t−1,t], is equal to the sum of the expected one period true re- turn, E [1 + r ], and a noise term stemming from microstructure frictions, σ2 . i,t−1,t ϕi,t−1 Therefore, the difference between equations (3.8) and (3.14) effectively captures the magnitude of the microstructure frictions as well as the effect of serially correlated

99 true returns. Specifically,

E [1 + rbi,t−2,t] E [1 + rbi,t−1,t] − E [1 + rbi,t−2,t−1] ρ σ2 ≈ E [1 + r ] + σ2 − E [1 + r ] − i i i,t−1,t ϕi,t−1 i,t−1,t 2 (1 − ρi ) E [1 + ri,t−2,t−1] ρ σ2 ≈ σ2 − i i ϕi,t−1 2 (1 − ρi ) (3.15)

where the last step relies on the assumption that for a short enough time period (a

day for our empirical analysis) the expected lagged one-period true return will be very

close to one (i.e. E [1 + ri,t−2,t−1] = 1). If we further assume that true returns are

independently distributed over time (i.e. ρi = 0), then the autocorrelation term goes

away and we are left with only the term related to microstructure frictions, σ2 .33 ϕi,t−1 As such, equation (3.15) provides the theoretic foundation of our microstructure fric-

tion measure.

To operationalize the above idea empirically, we define the sample analog of equa-

tion (3.15) as our measure of microstructure frictions, FRICi,t:

n=N Pn=N ! 1 X (1 + ri,t−n−1) (1 + ri,t−n) FRIC = (1 + r ) − n=1 b b (3.16) i,t N bi,t−n Pn=N n=1 n=1 (1 + rbi,t−n−1) The first term of equation (3.16) is the simple average of daily observed returns for firm i at time t computed using the previous N days of data, which is the sample

analog of E [1 + rbi,t−1,t]. The second term of equation (3.16) is the average two- day total return divided by the average lagged one-day return, which represents the

33If true returns are serially correlated, equation (3.15) will also pick up the extra autocorrelation term. In the case of negatively autocorrelated true returns (ρi < 0), equation (3.15) will overes- timate the amount of microstructure frictions. On the other hand, if true returns are positively autocorrleated (ρi > 0), equation (3.15) will underestimate the amount of frictions. We assume the baseline case of ρi = 0 for our main empirical analysis, but also present robustness evidence in Appendix C.1 where we explicitly account for the impact of the serial correlation in true returns. It is straightforward to also allow the microstructure noise term, ϕi,t, to be autocorrelated as in Brennan and Wang(2010) and Asparouhova, Bessembinder, and Kalcheva(2013).

100 sample analog of E [1 + rbi,t−2,t] /E [1 + rbi,t−2,t−1]. Note that the only required input to compute FRIC is daily stock returns.

The intuition for FRIC can be summarized as follows. The first term in equation

(3.16) is the simple average of daily observed returns, which, as an estimate of the true average daily return, has an upward bias that arises from microstructure frictions.

The second term produces a consistent friction-free estimate of the true average daily return by dividing the average two-day observed return by the average lagged one-day observed return, both of which share the same upward bias induced by microstructure frictions that is cancelled out by the division. Therefore, we can interpret FRIC, which is the difference between the two terms, as an estimate of the effect of microstructure frictions in average daily observed returns.

3.3 Data

Our study employs every listed security (AMEX/NYSE/Nasdaq) on the Center for

Research in Security Prices (CRSP) data files with share codes 10 or 11 from July,

1963 to June, 2013. We compute our friction measure, FRIC, following equation

(3.16) at the end of each month using past daily return data individual firms. We use previous 1 year of daily returns (approximately N = 250 daily observations per

firm) to estimate FRIC. However, we also conduct robustness checks by varying the number of daily observations, N, and our results are largely unchanged.

We obtain the return series of the market, size, value, and momentum factors from Ken French’s website, the Pastor and Stambaugh(2003) liquidity risk factor from CRSP, and the Hou, Xue, and Zhang(2015) q-factors from the authors. Hou,

Xue, and Zhang(2015, 2017) show that their q-factor model outperforms the Fama

101 and French(1993) three-factor and five-factor models and the Carhart(1997) four- factor model in explaining a broad set of asset pricing anomalies.

We are also interested in studying how our friction measure interacts with other

firm level characteristics to explain the cross section of stock returns. In particular, we want to ensure that our friction measure does not simply capture the effects of existing microstructure liquidity proxies. We construct firm level characteristics and liquidity variables as follows.

• Size – Following Fama and French(1992, 1993), size is defined as a firm’s market

equity at the end of June of year t.

• BE/ME – Following Fama and French(1993), book value of equity (BE) is

defined as book value of stockholder’s equity plus balance sheet deferred taxes

and investment credits minus the book value of preferred stock, for the fiscal

year ending in year t − 1. ME is market equity at the end of December of year

t − 1.

• IVOL – Following Ang, Hodrick, Xing, and Zhang(2006), idiosyncratic volatil-

ity (IVOL) is defined as the standard deviation of the residuals from regressing

a stock’s returns on the Fama and French(1993) three-factor model using the

previous month’s daily returns.

• Ret−1:−1, Ret−12:−2, and Ret−36:−13, – Previous month’s return, cumulative re-

turn over the past year (skipping the most recent month), and cumulative return

over the past three years (skipping the most recent year), respectively.

• Accrual – Prior to 1988, accruals are calculated using the balance sheet method

(Sloan(1996)) as changes in non-cash current assets, less changes in current

102 liabilities excluding changes in short-term debt and changes in taxes payable,

minus depreciation and amortization expenses. Starting in 1988, we use the cash

flow method (Hribar and Collins(2002)) to calculate accruals as the difference

between earnings and cash flows from operations.

• Profitability – Earnings (income before extraordinary items, minus dividends

on preferred stock, if available, plus income statement deferred taxes) for fiscal

year t − 1 divided by total assets at fiscal year-end of t − 2.

• Asset Growth – Total assets at fiscal year-end of t − 1 minus total assets at

fiscal year-end of t − 2 divided by total assets at fiscal year-end of t − 2.

• Volume – Average daily dollar volume from the previous year.

• BIDASK – Average daily effective bid-ask spread from the previous year, where

Ask−Bid daily effective bid-ask spread is computed as (Ask+Bid)/2 .

• Turnover – Average daily number of shares traded divided by number of shares

outstanding from the previous year.

• Amihud – Amihud(2002) illiquidity measure is computed as daily absolute

return divided by daily dollar trading volume, averaged over the previous year.

• Price – Share price at the end of the previous month.

• CS – Following Corwin and Schultz(2012), we first compute monthly spreads

as the averages of all daily high-low price spreads within each calendar month,

where we set negative spread estimates to zero, and require a minimum of 12

daily spreads to calculate a monthly spread. CS is then the 6-month average of

the monthly spreads.

103 • ZP – Lesmond, Ogden, and Trzcinka(1999) zero return proportion measure is

the number of days with zero return divided by number of trading days over

the previous year.

• C– Hasbrouck(2009) effective cost measure downloaded from the author’s

website.

Note that Volume, BIDASK, Turnover, Amihud, Price, CS, ZP, and C can all be viewed as alternative microstructure liquidity proxies. In the following sections, we show that our friction measure conveys useful pricing information above and beyond these conventional proxies.

3.4 Characteristics of Friction-Sorted Portfolios

Before discussing the impact of our friction measure on the cross section of stock returns, it is useful to examine which types of firms experience significant microstruc- ture frictions. At the beginning of each month, we use daily return data from the previous year to estimate our friction measure, FRIC, and then sort firms into 10 decile portfolios based on this measure. We require firms to have at least 50 non- missing daily returns from the previous year to be included in the analysis. Table 3.1 reports the average characteristics of the 10 FRIC-sorted portfolios over the time pe- riod from July, 1963 to June, 2013. Of particular interest to us are firms in Decile 10, the portfolio with the largest microstructure frictions. We also report the t-statistics

for the differences in average characteristics between Decile 10 and Decile 1 as well

as the time-series averages of the cross-sectional Pearson and Spearman correlations

between FRIC and other characteristics.

Table 3.1 shows that the average FRIC for Decile 10 is significantly larger than

104 Table 3.1: Characteristics of Friction-Sorted Portfolios At the beginning of each month, stocks are ranked by their friction measure (FRIC) and sorted into deciles. Equal-weighted average charac- teristics of these decile portfolios are computed over the following month from July, 1966 to June, 2013. Average characteristics of the friction portfolios are reported for the FRIC measure, average daily return, FRIC nzr nzv (FRIC constructed using non-zero return and non-zero volume days), FRIC nzv (FRIC constructed using non-zero volume days), size (market capitalization in $ millions), book-to-market equity ratio (BE/ME), idiosyncratic volatility (IVOL: standard deviation of residuals from Fama and French(1993) three-factor model using daily returns from previous month), market beta (slope coefficient on market factor from regression of each stock’s return on Fama and French(1993) three-factor model using daily returns from previous month), prior month’s return (month t − 1, Ret−1:−1 (%)), cumulative average return over 105 the past year skipping the most recent month (from month t − 12 to t − 2, Ret−12:−2 (%)), cumulative average return over the past three years skipping the most recent year (from month t − 36 to t − 13, Ret−36:−13 (%)), accrual (Prior to 1988, accruals are calculated using the balance sheet method as changes in non-cash current assets, less changes in current liabilities excluding changes in short-term debt and changes in taxes payable, minus depreciation and amortization expenses. Starting in 1988, accruals are calculated using the cash flow statement method as the difference between earnings and cash flows from operations), profitability (equity income divided by total assets), asset growth (change in total assets divided by total assets), average daily dollar trading volume from previous year ($ millions), average daily effective bid-ask spread, average daily turnover (average daily number of shares traded, divided by number of shares outstanding), Amihud(2002) illiquidity measure (daily absolute return divided by daily dollar trading volume, averaged over prior year) ×105, average share price, average number of trading days per year, Corwin and Schultz(2012) spread estimate (CS), zero return proportion measure (ZP: number of days with zero return divided by number of trading days from previous year), and Hasbrouck(2009) effective cost measure (C). Also reported are t-statistics for testing the equality of these characteristics between friction deciles 1 and 10. The final two columns report the time-series averages of the cross-sectional Pearson and rank correlations of each firm characteristic with the FRIC measure. Table 3.1 - Continued. FRIC-Sorted decile portfolios, July 1966 to December 2013 t-statistics Correlation with FRIC 1 2 3 4 5 6 7 8 9 10 (10-1) Pearson Rank Firm Characteristics FRIC (%) -0.03% -0.01% -0.01% 0.00% 0.00% 0.01% 0.01% 0.02% 0.04% 0.16% 30.06 Average Daily Return (%) 0.11% 0.08% 0.07% 0.06% 0.06% 0.06% 0.06% 0.07% 0.08% 0.19% 8.55 0.23 0.04 FRIC nzr nzv (%) -0.03% -0.01% -0.01% 0.00% 0.00% 0.01% 0.01% 0.02% 0.05% 0.20% 25.21 0.97 0.99 FRIC nzv (%) -0.03% -0.01% -0.01% 0.00% 0.00% 0.00% 0.01% 0.02% 0.04% 0.14% 33.35 0.93 0.95 Size (millions $) 653 1,836 2,698 2,862 2,604 2,073 1,391 724 309 80 -21.04 -0.05 -0.35 BE/ME 0.95 0.83 0.84 0.85 0.87 0.90 0.96 1.06 1.22 1.71 14.15 0.15 0.18 IVOL (%) 3.03% 2.15% 1.89% 1.81% 1.86% 2.03% 2.29% 2.72% 3.44% 5.60% 23.04 0.42 0.27 Beta 1.19 1.05 0.98 0.94 0.93 0.92 0.91 0.88 0.86 0.77 -17.73 -0.04 -0.06 Ret−1:−1 (%) 2.10% 1.41% 1.25% 1.19% 1.15% 1.12% 1.04% 1.08% 1.05% 1.50% -1.15 -0.01 -0.03

106 Ret−12:−2 (%) 32.62% 20.40% 16.99% 15.10% 14.34% 13.22% 11.79% 10.45% 7.93% 1.26% -13.31 -0.12 -0.16 Ret−36:−13 (%) 61.16% 50.94% 43.95% 40.74% 39.00% 38.70% 36.14% 32.32% 24.01% 5.24% -16.64 -0.12 -0.19 Accrual (%) -3.05% -2.89% -2.90% -2.96% -2.94% -2.95% -2.64% -2.95% -2.95% -4.93% -6.3 -0.04 -0.02 Profitability (%) 0.13% 5.16% 6.00% 6.17% 5.97% 5.32% 4.42% 2.81% 0.72% -4.81% -10.42 -0.12 -0.18 Asset Growth (%) 26.21% 19.98% 16.94% 15.81% 15.79% 15.98% 18.05% 16.76% 15.71% 10.40% -19.48 -0.06 -0.13

Liquidity Variables Volume (millions $) 5.51 10.36 12.85 13.21 12.29 10.32 7.82 4.82 2.23 0.56 -14.4 -0.07 -0.37 BIDASK (%) 2.77% 2.02% 1.87% 1.86% 1.91% 2.11% 2.50% 3.16% 4.27% 7.65% 21.23 0.47 0.36 Turnover (%) 4.73% 3.66% 3.22% 3.04% 2.94% 2.87% 2.72% 2.47% 2.17% 1.93% -21.84 -0.14 -0.27 Amihud 0.32 0.12 0.13 0.18 0.12 0.19 0.28 0.51 1.17 6.79 18.78 0.49 0.39 Average price ($) 19 41 49 46 47 41 30 19 13 6 -30.26 -0.05 -0.37 Number of trading days 244 246 245 245 244 243 240 236 230 220 -26.44 -0.17 -0.27 CS (%) 1.54% 0.99% 0.90% 0.92% 0.96% 1.12% 1.40% 1.83% 2.57% 5.05% 23.81 0.48 0.39 ZP (%) 15.04% 13.55% 13.64% 14.09% 14.76% 15.83% 17.40% 19.35% 21.78% 26.67% 18.78 0.31 0.37 C ×100 0.71 0.53 0.49 0.50 0.56 0.67 0.88 1.19 1.70 3.16 30.02 0.61 0.49 that for Decile 1. In addition, firms with larger frictions are small, value, volatile firms with poor recent performance (in terms of both returns and accounting profitability).

Table 3.1 also reports the average values for a number of traditional microstructure liquidity proxies. These variables are Volume, BIDASK, Turnover, Share Price, Num- ber of Trading Days, Amihud, CS, ZP, and C. Not surprisingly, high FRIC firms also have significantly lower dollar trading volume, larger bid-ask spreads, lower share turnover, lower share prices, and smaller number of trading days.

The last two columns of Table 3.1 show that FRIC is strongly correlated with the above microstructure liquidity proxies. The rank correlations between FRIC and

BIDASK, Amihud, CS, ZP, and C are 0.36, 0.39, 0.39, 0.37, and 0.49, respectively. In addition, FRIC is negatively correlated with Turnover, Share Price, and the Number of Trading Days with rank correlations of -0.27, -0.37, and -0.27, respectively. Given these correlations, it is important to control for these liquidity variables when exam- ining the return predictability of FRIC.

In Figure 3.1, we plot the market-wide average FRIC each year from 1966 to

2013. For each firm, the monthly FRIC values are averaged within each year. Then, annual firm-level FRIC values are averaged across all firms. In Figure 3.1a, we plot the average FRIC value of NYSE/AMEX and Nasdaq firms separately, and observe that they show considerable variation over time. While FRIC remains relatively low in the late 1960s through the early 1970s, it spikes during the recessions of 1974 to

1975 and 1990 to 1991, the tech bubble in the late 1990s and subsequent crash in the early 2000s (for Nasdaq firms in particular), and the financial crisis of 2007 to

2009. In Figures 3.1b and 3.1c, we plot the average FRIC value for different price

107 Figure 3.1a. Average FRIC Estimates from 1966-2013 (NYSE/AMEX & NASDAQ) 3.5 17.5 1990-1991 recession 3 15 2007-2009 Financial Crisis 2.5 12.5

2 Tech Bubble 10 1974-1975 recession FRIC NYSE/AMEX 1.5 7.5 (basis points) NASDAQ 1 5

0.5 2.5

0 0

108 -0.5 -2.5

Figure 3.1b. Average FRIC Estimates from 1966-2013 (Price Groups) 60

50

40

30 Price<1 FRIC (basis points) $1$5

10

0

-10

Figure 3.1c. Average FRIC Estimates from 1966-2013 (Market Cap Groups) 16

14

12

10

8 Micro Cap FRIC (basis points) Mid Cap 6 Large Cap 4

2 109 0

-2

Figure 3.1: Time-Series of Market Average Friction Estimates

The annual market-wide cross-sectional average of our friction measure, FRIC, is plotted for the period 1966 to 2013. For each stock, the friction measure is averaged within each year. Then, annual firm-level friction estimates are averaged across all firms. Only firms with at least 50 daily non-missing observations within the year are included in the sample each year. Figure 3.1a plots the equal weighted average FRIC each year for stocks listed on NYSE/AMEX and NASDAQ separately. Figure 3.1b plots the equal weighted average FRIC each year for stocks within different price groups (prior year’s share price smaller than $1, between $1 and $5, and larger than $5) separately. Figure 3.1c plots the equal weighted average FRIC each year for micro-cap (below 20th NYSE percentile), mid (small)-cap (between 20th and 50th NYSE percentile), and large-cap (above 50th NYSE percentile) stocks separately. and market capitalization groups. They show that the spikes in FRIC are concen- trated among low price and microcap firms. In sum, these figures suggest that FRIC captures important market-wide liquidity events. We also note that FRIC not only allows us to study market-wide microstructure frictions over a much longer time pe- riod, but it also exhibits more time-series variation than various liquidity and spread measures reported by Angel, Harris, and Spatt(2011, 2015).

Our baseline FRIC construction uses all available CRSP daily returns from the previous year, which are based on the average of bid and ask prices when the closing price is missing. This raises a possibility that FRIC could potentially underestimate the severity of microstructure frictions since averaging bid and ask prices will smooth out the noise introduced by microstructure effects. To address this issue, we construct two variations of FRIC: FRIC nzr nzv, which uses daily returns from the previous year excluding days with both zero returns and zero volume, and FRIC nzv, which excludes days with zero volume. Table 3.1 shows that the magnitudes of the different friction measures are very similar for each FRIC decile portfolio. For example, FRIC

(our main friction measure), FRIC nzr nzv, and FRIC nzv are all -0.03% for Decile

1 and are 0.16%, 0.20%, and 0.14%, respectively, for Decile 10. The rank correlations between FRIC and the two alternative measures are 0.99 and 0.95, respectively. Most importantly, the main results of our paper are robust to using the two alternative fric- tion measures, suggesting that the inclusion of non-traded prices and zero-return days does not bias the inferences of our analysis.

110 3.5 Cross-Sectional Return Predictability: Portfolio Sorts

Table 3.2 reports the average returns of decile portfolios sorted on FRIC. At the beginning of each month, stocks are ranked by FRIC calculated using the previous year’s daily return data and sorted into deciles. Equal- and value-weighted monthly returns of the decile portfolios are computed for the current month, and the deciles are rebalanced at the beginning of next month.

3.5.1 Returns of FRIC Portfolios

Panel A of Table 3.2 shows that both the equal-weighted and value-weighted average returns increase with FRIC. For example, the equal-weighted average return increases monotonically from 0.78% per month for Decile 1 to 2.05% per month for

Decile 10. The average return spread between Deciles 10 and 1 is a striking 1.27%

(t = 7.00) per month for equal-weighted returns and 0.82% (t = 3.49) per month for value-weighted returns, both of which are economically large and statistically highly significant.

To account for differences in risk exposures between Deciles 10 and 1, the last four columns regress the monthly 10-1 return spreads on the Fama and French(1993) three-factor model, the Carhart(1997) four-factor model, a five-factor model that adds the Pastor and Stambaugh(2003) liquidity factor to the Carhart model, and the

Hou, Xue, and Zhang(2015) q-factor model. All the alphas are large and significant, ranging from 1.16% to 1.32% per month for equal-weighted returns and 0.58% to

0.76% per month for value-weighted returns. Thus, the return spreads between FRIC

Deciles 10 and 1 are not driven by differences in their exposures to common risk factors.

111 Table 3.2: Microstructure Frictions and the Cross-Section of Expected Stock Returns: Portfolio Sorts The equal- and value-weighted average monthly returns (%) of FRIC-sorted decile portfolios, their t-statistics (in parentheses), and the difference in returns between decile portfolios 10 (largest FRIC) and 1 (smallest FRIC) are reported over the period July, 1966, to June, 2013. Panel A reports the equal- and value-weighted results using the baseline FRIC measure over the full sample. The last four columns in Panel A report the intercepts, or alphas, from time-series regressions of the 10-1 average return spreads on the Fama and French(1993) three-factor model, Carhart (1997) four-factor model (which adds a momentum factor to Fama and French(1993) factors), a five-factor model that adds the Pastor and Stambaugh(2003) aggregate liquidity risk factor-mimicking portfolio to the Carhart(1997) four-factor model, and the Hou, Xue, and Zhang (2015), in short HXZ, q-factor model. Panel B reports the equal- and value-weighted return spreads between FRIC decile portfolios 10 and 1 as well as their HXZ q-factor alphas using 3-day vs. lagged 2-day and 4-day vs. lagged 3-day average returns to estimate friction-free average daily returns, different time horizons of past daily returns (6-month, 3-year, and 5-year) to compute average daily returns for the friction measure, for NYSE/AMEX and NASDAQ stocks separately, for firms with share prices above $1 and $5, for firms excluding micro-cap (defined as firms below NYSE 20th and 50th percentile market capitalization), for two sub-periods, and for January and non-January months separately.

Panel A: Returns on FRIC-sorted portfolios (10-1) (10-1) (10-1) (10-1) Decile 1 2 3 4 5 6 7 8 9 10 (10-1) FF Carhart Pastor and HXZ

112 portfolio 3-factor 4-factor Stambaugh q-factor 5-factor Equal-weighted 0.78 1.04 1.12 1.20 1.23 1.25 1.27 1.37 1.41 2.05 1.27 1.16 1.22 1.21 1.32 (2.26) (3.88) (4.80) (5.53) (5.86) (5.77) (5.60) (5.49) (4.94) (5.57) (7.00) (6.86) (7.08) (6.81) (7.14) Value-weighted 0.47 0.89 0.87 0.96 0.94 0.91 0.97 0.98 1.02 1.29 0.82 0.58 0.73 0.76 0.76 (1.39) (3.60) (4.24) (4.92) (5.01) (4.61) (4.45) (3.96) (3.50) (3.58) (3.49) (2.68) (3.32) (3.34) (3.30)

Panel B: Robustness. 10-1 return spreads and HXZ q-factor alphas 3-day vs. 4-day vs. 6-month 3-year 5-year NYSE/ lagged 2-day lagged 3-day horizon horizon horizon AMEX NASDAQ (10-1) Alpha Equal-weighted 1.12 1.28 1.08 1.29 0.99 1.17 1.09 1.02 0.80 0.70 0.92 1.08 1.41 1.47 (6.24) (6.89) (5.74) (6.61) (5.78) (6.45) (6.61) (6.22) (5.51) (4.77) (4.77) (5.56) (6.28) (6.53) Value-weighted 0.56 0.80 0.46 0.76 0.41 0.57 0.84 0.78 0.57 0.35 0.53 0.47 0.88 0.90 (2.11) (3.00) (1.75) (2.98) (1.73) (2.30) (3.56) (3.56) (2.59) (1.75) (2.44) (2.09) (2.94) (2.98)

Price Price Exclude 07/66 to 01/91 to ≥ $1 ≥ $5 Micro cap 12/90 06/13 Jan Feb-Dec (10-1) Alpha Equal-weighted 0.70 0.55 0.34 0.13 0.41 0.19 0.78 0.87 1.80 2.16 7.73 6.32 0.68 0.58 (4.48) (3.67) (2.58) (1.10) (3.12) (1.59) (3.93) (4.60) (5.85) (7.34) (9.08) (6.64) (4.29) (3.61) Value-weighted 0.57 0.50 0.21 0.18 0.39 0.31 0.56 0.53 1.11 1.29 6.28 4.27 0.33 0.33 (2.79) (2.57) (1.38) (1.14) (2.40) (1.88) (2.16) (1.97) (2.76) (3.38) (4.91) (2.90) (1.51) (1.51) 3.5.2 Robustness

Our results are robust to alternative methods to estimate FRIC and to various

subperiod and subsample analysis. We discuss these robustness tests in the following

subsections.

Alternative FRIC Constructions

Our baseline FRIC measure is based on obtaining a friction-free average daily

return estimate by dividing the average two-day observed return by the average lagged

one-day observed return. This is a natural choice which allows us the maximum

amount of data within a given window to estimate FRIC. However, this specification

has a potential drawback for illiquid stocks. These stocks are more likely to have

no trades, and thus no change in prices, over two days in a row. As a result, FRIC

could be downward biased for these stocks. To investigate the extent of this potential

bias and more in general the robustness of our baseline specification, we construct

two alternative versions of FRIC where the friction-free average daily return estimate

is obtained as the average three-day observed return divided by the average lagged

two-day observed return and the average four-day observed return divided by the

average lagged three-day observed return, both of which are estimated using previous

one year’s data. The first four columns of the top part of Table 3.2 Panel B report

that the average equal-weighted 10-1 monthly return spread associated with these

two alternative FRIC measures are 1.12% (t = 6.24) with a q-factor alpha of 1.28%

(t = 6.89) and 1.08% (t = 5.74) with a q-factor alpha of 1.29% (t = 6.61), respectively.

The value-weighted q-factor alphas are 0.80% (t = 3.00) and 0.76% (t = 2.98) per month, respectively. Therefore, our results are robust to alternative specifications of

113 the friction-free average daily return estimate.

We also check the robustness of our results in terms of the amount of past daily return data we use to compute FRIC. In our baseline specification, we use the previous year’s daily returns. The next six columns of the top part of Panel B report the average 10-1 return spreads and q-factor alphas for alternatively constructed FRIC measures using previous six months, three years, and five years of daily returns. The equal-weighted 10-1 return spreads are 0.99% (t = 5.78), 1.09% (t = 6.61), and 0.80%

(t = 5.51), respectively, and the corresponding q-factor alphas are 1.17% (t = 6.45),

1.02% (t = 6.22), and 0.70% (t = 4.77), respectively, which are all economically large and statistically highly significant. Similar to the baseline results, the value-weighted

10-1 return spreads and q-factor alphas are smaller in magnitude than their equal- weighted counterparts. Overall, our results are robust to using different sample size of daily returns to estimate the FRIC measure.

Subsamples

The last four columns of the top part of Panel B report the average 10-1 return spreads and q-factor alphas of FRIC-sorted portfolios for NYSE/AMEX and Nasdaq

firms separately. The return spreads and q-factor alphas are significant for both the

NYSE/AMEX and Nasdaq subsamples, yet of higher magnitudes for the Nasdaq sub- sample which is likely due to the fact that Nasdaq listings include a larger proportion of smaller and less liquid firms.

We then explore whether and how our results change after imposing a number of

filters commonly used in the literature for microstructure frictions such as price and market cap filters (see, e.g., Jegadeesh and Titman(2001), Bali and Cakici(2008),

114 Fama and French(2008)). The first four columns of the bottom part of Panel B re-

stricts the sample to firms with share price at the end of the previous month greater

than or equal to $1 and $5. The next two columns exclude firms with market capi-

talizations less than the 20th percentile based on NYSE breakpoints (following Fama and French(2008)). Not surprisingly, the magnitudes of the return spreads and q- factor alphas are reduced considerably, although they generally remain statistically significant (the exception being value-weighted return spreads and q-factor alphas when price ≥ $5). For example, removing firms with share prices less than $1 at the end of the previous month reduces the average equal-weighted return spread to 0.70%

(t = 4.48) and q-factor alpha to 0.55% (t = 3.67). This reduction is consistent with the pricing impact of microstructure frictions being more pronounced for low priced and small firms, which likely suffer the most severe frictions.

Subperiods

The last eight columns of the bottom part of Panel B report the average 10-1 return spreads and q-factor alphas for two subperiods (1966 to 1990 and 1991 to

2013), and for January and non-January months separately. The return spreads and alphas are significant in both subperiods, but higher in the second subperiod, which is likely due to the introduction of Nasdaq firms into our sample. In addition, both the return spreads and q-factor alphas are huge in January, with an average equal- weighted 10-1 spread of 7.73% (t = 9.08) and q-factor alpha of 6.32% (t = 6.64) per month! On the other hand, both the spreads and alphas are much smaller in

February-December, more so for value-weighted returns. Given that high FRIC firms are generally small firms, this is consistent with the well-documented January effect where the returns of small firms are much higher for the month of January, and the

115 bulk of the size anomaly is attributable to these abnormal January returns (see, e.g.,

Keim(1983), Roll(1983), Lakonishok and Smidt(1984)).

3.6 Cross-Sectional Return Predictability: Fama-MacBeth Regressions

We complement the portfolio sorts and time-series factor regressions with Fama

and MacBeth(1973) cross-sectional regressions in Table 3.3. These regressions pro-

vide additional robustness checks for our results since they employ all firms without

imposing portfolio breakpoints, allow for more control variables including traditional

liquidity and microstructure proxies, and provide an alternative weighting scheme for

portfolios.34

Each month, we regress the cross section of individual stock returns (in excess of

the one-month T-bill rate) on FRIC and other firm characteristics. The firm charac-

teristics include the logarithm of size (market cap), logarithm of BE/ME, previous

month’s stock return (month t − 1), previous year’s stock return skipping the most

recent month (from month t12 to t2), previous three year’s stock return skipping

the most recent year (from month t36 to t13), and the logarithm of prior month end’s share price. In addition, we include Profitability and Asset Growth to control for the well-known results that more profitable firms and firms that invest less have higher average returns (See, e.g., Haugen and Baker(1996), Cooper, Gulen, and Schill

(2008), and Hou, Xue, and Zhang(2015)). Furthermore, we use Accruals and IVOL to control for the negative return predictability of operating accruals (Sloan(1996))

34Each coefficient from a Fama-MacBeth regression is the return to a minimum variance portfolio with weights that sum to zero, weighted characteristic on its corresponding regressor that sums to one, and weighted characteristics on all other regressors that sum to zero. The weights are tilted toward firms with the most extreme (volatile) returns (see Hou and Moskowitz(2005)).

116 Table 3.3: Microstructure Frictions and the Cross-Section of Expected Stock Returns: Fama-Macbeth Regressions

Panel A presents results from Fama and MacBeth(1973) monthly cross-sectional regressions of stock returns in excess of the one-month T-bill rate on our friction measure (FRIC), log of firm size (market capitalization), log of book-to-market equity (BE/ME), previous month’s return (month t − 1, Ret−1:−1 (%)), previous year’s return (from month t − 12 to t − 2, Ret−12:−2 (%)), previous three year’s return (from month t − 36 to t−13, Ret−36:−13 (%)), log of previous month’s share price, profitability (equity income divided by total assets), asset growth (change in total assets divided by total assets), accrual (Prior to 1988, accruals are calculated using the balance sheet method as changes in non-cash current 117 assets, less changes in current liabilities excluding changes in short-term debt and changes in taxes payable, minus depreciation and amortization expenses. Starting in 1988, accruals are calculated using the cash flow statement method as the difference between earnings and cash flows from operations), idiosyncratic volatility (IVOL: standard deviation of residuals from Fama and French(1993) three-factor model using daily returns from previous month), and a host of alternative liquidity variables over the period July, 1966, to June, 2013. Liquidity variables include average daily turnover (average daily number of shares traded, divided by number of shares outstanding from prior year), volume (average daily dollar trading volume over the past year), Amihud(2002) illiquidity measure (daily absolute return divided by daily dollar trading volume, averaged over prior year), average daily effective bid-ask spread, Corwin and Schultz(2012) spread estimate (CS), zero return proportion measure (ZP: number of days with zero return divided by number of trading days from previous year), and Hasbrouck(2009) effective cost measure (C), each defined separately for NYSE/AMEX and Nasdaq traded firms. The time-series averages of the coefficient estimates and their associated time-series t-statistics (in parentheses) are reported in the style of Fama and MacBeth(1973). Panel B shows results from two kitchen sink regressions, where all control variables and liquidity measures are included in the regressions. The first kitchen sink specification excludes FRIC and the second specification includes all variables. Table 3.3 - Continued. Panel A: Fama-MacBeth Regressions Panel B: Fama-MacBeth Regressions (Kitchen Sink) Liquidity Measures (1) (2) NYSE Nasdaq Turnover Volume Amihud BIDASK CS ZP C FRIC 4.08 /AMEX (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (4.68) log(Size) 0.02 0.02 -0.04 0.15 0.00 0.05 0.03 0.04 -0.01 0.00 0.01 Turnover 0.05 0.05 (0.69) (0.60) (-1.51) (2.19) (0.07) (0.84) (0.80) (1.40) (-0.17) (-0.11) (0.18) (0.26) (0.25) log(BE/ME) 0.24 0.23 0.20 0.28 0.21 0.21 0.21 0.20 0.20 0.22 0.21 Volume -0.06 -0.06 (4.51) (4.36) (3.96) (4.46) (4.38) (4.39) (4.24) (4.23) (3.83) (4.80) (3.93) (-0.32) (-0.29) Ret−1:−1 -6.03 -5.96 -5.25 -5.08 -6.15 -6.17 -6.00 -6.08 -6.03 -6.04 -6.40 Amihud 13.74 13.06 (-17.44) (-17.32) (-13.35) (-12.87) (-18.30) (-18.46) (-17.31) (-18.09) (-17.24) (-18.69) (-17.85) (1.67) (1.57) Ret−12:−2 0.64 0.64 0.63 0.50 0.67 0.68 0.65 0.67 0.70 0.65 0.73 BIDASK -16.05 -16.20 (5.24) (5.29) (4.07) (4.57) (5.61) (5.69) (5.30) (5.60) (5.63) (5.70) (5.64) (-2.89) (-2.91) Ret -0.10 -0.10 -0.08 -0.06 -0.10 -0.10 -0.10 -0.11 -0.10 -0.11 -0.10 CS -2.25 -4.41

−36:−13 NYSE/AMEX (-2.27) (-2.26) (-1.53) (-1.53) (-2.35) (-2.37) (-2.23) (-2.65) (-2.32) (-2.49) (-2.02) (-0.45) (-0.88)

118 Log(Price) -0.36 -0.33 -0.18 -0.48 -0.30 -0.30 -0.29 -0.36 -0.25 -0.33 -0.21 ZP -3.39 -3.32 (-4.29) (-3.98) (-2.48) (-3.58) (-3.76) (-3.60) (-3.50) (-4.08) (-3.00) (-3.50) (-2.23) (-4.30) (-4.22) Profitability 1.21 1.21 1.60 0.65 0.99 1.00 1.06 1.12 1.11 1.21 1.29 C 33.32 24.00 (4.81) (4.76) (5.25) (2.86) (3.89) (3.92) (3.98) (4.66) (4.15) (4.83) (4.43) (3.50) (2.49) Asset growth -0.24 -0.23 -0.26 -0.27 -0.21 -0.21 -0.22 -0.22 -0.22 -0.23 -0.26 Turnover -0.17 -0.15 (-4.77) (-4.58) (-4.01) (-5.30) (-3.71) (-3.66) (-3.72) (-4.41) (-3.88) (-4.78) (-3.98) (-1.17) (-1.07) Accruals -0.93 -0.90 -1.15 -0.55 -0.94 -0.94 -0.94 -0.85 -1.00 -0.90 -1.08 Volume 0.13 0.11 (-6.54) (-6.37) (-5.52) (-2.71) (-6.36) (-6.34) (-6.27) (-6.13) (-6.53) (-6.47) (-6.79) (0.87) (0.77) IVOL -6.46 -9.06 -15.86 -5.10 -10.63 -10.43 -13.65 -5.78 -14.43 -8.64 -14.72 Amihud 147.47 149.37 (-2.45) (-3.32) (-4.91) (-1.63) (-4.68) (-4.65) (-4.95) (-2.44) (-5.29) (-3.68) (-5.06) (1.46) (1.47) FRIC 4.55 7.02 2.56 4.93 4.93 4.31 4.61 5.70 4.51 4.54 BIDASK -10.13 -11.96 (4.85) (5.66) (3.32) (5.53) (5.59) (4.89) (4.85) (6.49) (5.15) (4.63) (-1.59) (-1.84) Nasdaq -0.13 -0.31 -0.79 -0.82 -0.79 -1.09 -0.90 NASDAQ CS -3.09 -3.26 (-0.25) (-0.12) (-2.31) (-0.87) (-2.01) (-1.56) (-2.25) (-0.43) (-0.46) Liquidity -0.07 -0.05 32.45 -8.60 -2.66 -1.30 15.74 ZP 3.67 3.92 AMEX/NYSE (-1.26) (-0.86) (3.75) (-2.03) (-0.61) (-1.53) (2.18) (0.85) (0.91) Liquidity -1.11 -0.04 43.24 14.29 8.89 1.98 37.21 C 67.85 60.10 Nasdaq (-1.48) (-0.15) (0.34) (0.67) (1.23) (1.06) (1.85) (1.10) (0.98) Control Variables Yes Yes and idiosyncratic volatility (Ang, Hodrick, Xing, and Zhang(2006)). We report the time-series averages of the cross-sectional regression coefficients along with their t- statistics in Table 3.3.

The first column in Panel A of Table 3.3 confirms the standard results found in the literature. Average stock returns are positively related to BE/ME, past one- year’s return (momentum), and profitability, while negatively related to past three- year’s return (long-term reversal), past one-month’s return (short-term reversal), As- set Growth, Accruals, Share Price, and IVOL. The second column adds our friction measure, FRIC, to the regression. FRIC is strongly positively associated with average returns with an average coefficient of 4.55 (t = 4.85), consistent with the portfolio sort results from Table 3.2. In terms of economic significance, Table 3.1 shows that the difference in average FRIC between Deciles 10 and 1 is approximately 19 basis points. Therefore, the coefficient on FRIC in Table 3.3 implies a difference in aver- age returns of 86 basis points per month between the two extreme deciles, which is economically sizable. This implied return difference is almost identical to the value- weighted 10-1 return spread (82 basis points) reported in Table 3.2, but smaller than the equal-weighted return spread (127 basis points).

One potential concern of the Fama and MacBeth(1973) regressions is that the re- sults may be driven by small and volatile Nasdaq firms which are given more weight due to least squares minimization. To address this concern, specifications (3) and

(4) of Panel A split the sample into NYSE/AMEX and Nasdaq subsamples and es- timate the regressions separately. The coefficient on FRIC is much larger among

NYSE/AMEX firms (7.02, t = 5.66), but it is also highly significant among Nasdaq

firms (2.56, t = 3.32).

119 To ensure that the return predictability associated with our FRIC measure is not simply a manifestation of previously documented liquidity effects, the next seven columns of Table 3.3 report results with various liquidity proxies included in the

Fama-MacBeth regressions. We control for seven liquidity proxies (one at a time) commonly used in the literature: Turnover, Volume, Amihud, BIDASK, CS, ZP, and

C. Since some of these variables rely on volume information and the reported volume on Nasdaq include interdealer trades while those on NYSE/AMEX do not, we com- pute these variables separately for NYSE/AMEX and Nasdaq firms and also include a Nasdaq exchange indicator in the regressions.

The results show that the FRIC effect is robust to the inclusion of traditional microstructure liquidity proxies. Despite the high correlations between FRIC and these variables, the economic magnitude and statistical significance of the coefficient on FRIC is virtually unchanged, and in some cases even larger and more significant.

Interestingly, many of the conventional liquidity proxies are no longer associated with a reliable premium once FRIC is controlled for. For example, the coefficients on turnover and volume are statistically insignificant. The coefficient on Amihud is only significant for NYSE/AMEX firms but not for Nasdaq firms. For BIDASK, CS, and

ZP, their coefficients have the wrong signs for NYSE/AMEX firms, suggesting that less liquid NYSE/AMEX firms have actually lower average returns. The coefficients for Nasdaq firms, though in the right direction, are all insignificant. In Panel B, we include all of the above liquidity proxies into one regression along with FRIC. The average coefficient on FRIC in this “kitchen sink” regression is 4.08 (t = 4.68), which is only slightly smaller than the coefficient when none of the traditional liquidity variables is included (4.55, t = 4.85). In this kitchen sink regression, most of the

120 traditional liquidity proxies exhibit similar weak results as before.

Another concern with the Fama-MacBeth regressions is that they employ real- ized stock returns, which are affected by the very same microstructure frictions our

FRIC measure is trying to capture, and thus their relation with FRIC could poten- tially be mechanical. We address this concern in three ways. First, recall that our

FRIC measure is calculated based on a rolling one-year window of past daily returns

(not the daily returns in the current month) and thus is not mechanically related to the monthly returns it predicts.35 Second, we follow Asparouhova, Bessembinder, and Kalcheva(2010, 2013) and use an alternative weighting procedure in the Fama-

MacBeth regressions where each return observation is weighted by the lagged total return on the same firm. The results are tabulated in Table C.2. Consistent with

Asparouhova, Bessembinder, and Kalcheva(2010, 2013), the effects associated with most of the conventional liquidity variables are weaker under the alternative weight- ing scheme. The effect of FRIC is also slightly weaker, but remains economically large and statistically highly significant. For example, the FRIC coefficient in the kitchen sink regression is 4.01 (t = 4.59), compared with 4.08 (t = 4.68) reported in

Table 3.3. As a final attempt to alleviate concerns about a mechanical relationship, in untabulated results, we include the FRIC measure calculated using current month’s daily returns on the right hand side of our Fama-MacBeth regressions and find that our main FRIC measure (based on past year’s daily returns) remains positive and highly significant, even after using the alternative weighting scheme of Asparouhova,

Bessembinder, and Kalcheva(2010, 2013).

35We also check the correlation between our FRIC measure (based on past one year’s daily returns) and the FRIC measure constructed using current month’s daily returns and find it to be very low. For example, the time-series average of the cross-sectional Pearson correlation is 0.19 and the cross- sectional average of the within-firm time-series Pearson correlation is 0.04.

121 3.7 Interactions between FRIC and Firm Characteristics

Although our analysis so far indicates that the return predictability of FRIC is not subsumed by a broad set of firm characteristics known to explain average returns, it is interesting to study how the pricing impact of FRIC varies with those firm char- acteristics.

Table 3.4 reports the FRIC premium, defined as the difference in returns between the highest and lowest quintile of FRIC-sorted portfolios, within quintiles first sorted on each firm characteristic. By examining the FRIC premium within each charac- teristic quintile, we are essentially controlling for these firm characteristics, while at the same time highlighting their interactions with FRIC by observing differences in the FRIC premium across characteristic quintiles. Table 3.4 shows that the FRIC premium mostly resides among small, value, volatile, past loser, low accruals, low profitability, and low asset growth firms, consistent with those firms being subject to more microstructure frictions. For example, the equal-weighted FRIC premium is 1.42% (t = 6.63) per month in the smallest Size quintile compared to only 0.17%

(t = 1.70) in the biggest Size quintile; 2.14% (t = 8.76) in the highest IVOL quin- tile compared to a mere 0.07% (t = 1.02) in the lowest IVOL quintile; and 2.01%

(t = 8.07) in the lowest Ret−12:−2 quintile compared to 0.50% (t = 3.98) in the high- est Ret−12:−2 quintile. The differences in the FRIC premium between the extreme characteristic quintiles are all highly significant. We observe similar patterns across characteristic quintiles when we use value-weighted returns.

We also investigate how the FRIC premium varies with other commonly used microstructure liquidity proxies. Consistent with the intuition that FRIC captures microstructure induced illiquidity, the FRIC premium is strongest among firms with

122 Table 3.4: Friction Premiums across Characteristic Quintiles Equal- and value-weighted average returns for double-sorted portfolios on various firm characteristics and FRIC are reported for the period July, 1966, to June, 2013. Friction premium is defined as the difference in average returns between stocks in the highest and lowest FRIC quintiles, within quintiles formed by first sorting on firm characteristics such as size (market capitalization in $ millions), book-to-market equity (BE/ME), idiosyncratic volatility (IVOL: standard deviation of residuals from Fama and French(1993) three-factor model using daily returns from previous month), prior month’s return (month t − 1, Ret−1:−1 (%)), cumulative average return over the past year skipping the most recent month (from month t − 12 to t − 2, Ret−12:−2 (%)), cumulative average return over the past three years skipping the most recent year (from month t − 36 to t − 13, Ret−36:−13 (%)), accrual (Prior to 1988, accruals are calculated using the balance sheet method as changes in non-cash 123 current assets, less changes in current liabilities excluding changes in short-term debt and changes in taxes payable, minus depreciation and amortization expenses. Starting in 1988, accruals are calculated using the cash flow statement method as the difference between earnings and cash flows from operations), profitability (equity income divided by total assets), asset growth (change in total assets divided by total assets), average share price, Amihud(2002) illiquidity measure (daily absolute return divided by daily dollar trading volume, averaged over prior year), average daily turnover (average daily number of shares traded, divided by number of shares outstanding), average daily effective bid-ask spread, average daily dollar trading volume from previous year ($ millions), zero return proportion measure (ZP: number of days with zero return divided by number of trading days from previous year), Corwin and Schultz(2012) spread estimate (CS), and Hasbrouck(2009) effective cost measure (C). At the end of each month, stocks are ranked by each of these characteristics and sorted into quintiles. Within each characteristic quintile, stocks are then sorted into quintiles based on their individual FRIC measure. The equal- and value-weighted average monthly return differences (% per month) between friction quintiles 5 and 1, or friction premiums, and their t-statistics (in parentheses) are reported within each characteristic quintile. Finally, the friction premium spreads between the high and low characteristic quintiles and their associated t-statistics are also reported. Table 3.4 - Continued. Equal-weight Value-weight Low 2 3 4 High High - Low Low 2 3 4 High High - Low Firm Characteristics Size 1.42% 0.64% 0.40% 0.25% 0.17% -1.25% 0.99% 0.62% 0.38% 0.21% 0.08% -0.90% (6.63) (4.45) (3.50) (2.23) (1.70) (-5.33) (4.85) (4.38) (3.23) (1.80) (0.72) (-3.91) BE/ME 0.50% 0.20% 0.24% 0.32% 1.27% 0.77% 0.26% -0.16% 0.02% 0.14% 0.89% 0.63% (3.30) (1.73) (2.06) (2.42) (6.06) (3.79) (1.22) (-0.93) (0.11) (0.77) (2.85) (1.72) IVOL 0.07% 0.15% 0.37% 0.67% 2.14% 2.07% 0.05% 0.15% 0.42% 0.61% 2.28% 2.23% (1.02) (1.77) (3.15) (4.44) (8.76) (8.54) (0.43) (1.10) (2.46) (2.90) (5.96) (5.67) Ret−1:1 2.16% 0.41% 0.28% 0.36% -0.07% -2.22% 1.52% 0.08% -0.15% -0.12% -0.20% -1.71% (9.49) (3.11) (2.44) (2.90) (-0.46) (-10.65) (4.70) (0.42) (-0.81) (-0.71) (-1.01) (-5.16) Ret−12:2 2.01% 0.74% 0.63% 0.44% 0.50% -1.50% 1.67% 0.35% 0.10% 0.04% 0.11% -1.56% (8.07) (5.04) (5.84) (4.16) (3.98) (-6.12) (4.46) (1.80) (0.61) (0.26) (0.66) (-3.92) Ret−36:−13 1.12% 0.54% 0.27% 0.32% 0.29% -0.84% 0.85% 0.26% 0.03% 0.18% 0.20% -0.65% (5.20) (3.67) (2.43) (2.88) (2.42) (-4.12) (2.84) (1.09) (0.15) (1.04) (1.17) (-1.99) Accrual 1.18% 0.65% 0.71% 0.66% 0.59% -0.59% 0.74% 0.40% 0.33% 0.12% 0.35% -0.39% (5.98) (4.19) (5.13) (4.13) (3.72) (-3.76) (2.90) (1.64) (1.73) (0.51) (1.45) (-1.35) Profitability 1.52% 0.62% 0.37% 0.31% 0.31% -1.20% 1.15% 0.31% 0.02% 0.04% 0.38% -0.76% 124 (6.98) (3.76) (3.11) (2.83) (2.74) (-6.22) (3.77) (1.17) (0.10) (0.22) (2.12) (-2.37) Asset Growth 1.26% 0.56% 0.33% 0.24% 0.46% -0.80% 0.74% 0.30% 0.23% 0.02% 0.25% -0.48% (6.17) (3.84) (2.77) (1.89) (2.97) (-4.33) (2.68) (1.39) (1.28) (0.10) (1.23) (-1.60) Liquidity Variables Share Price 1.55% 0.45% 0.27% 0.03% -0.08% -1.63% 1.38% 0.60% 0.28% 0.32% -0.11% -1.48% (7.12) (3.28) (2.21) (0.30) (-0.80) (-6.89) (3.75) (2.53) (1.72) (2.14) (-0.92) (-3.82) Amihud 0.15% 0.15% 0.36% 0.46% 1.30% 1.15% 0.03% 0.19% 0.41% 0.44% 0.78% 0.75% (1.40) (1.33) (3.22) (3.17) (5.68) (4.41) (0.28) (1.50) (3.40) (2.85) (3.23) (2.73) Turnover 0.92% 1.08% 0.80% 0.43% 0.46% -0.45% 0.39% 0.58% 0.28% 0.13% 0.38% -0.02% (5.12) (5.82) (4.17) (2.71) (3.26) (-2.38) (1.81) (2.59) (1.23) (0.73) (1.99) (-0.06) BIDASK 0.18% 0.41% 0.50% 0.67% 1.43% 1.25% 0.16% 0.30% 0.18% 0.52% 1.20% 1.04% (1.54) (4.09) (3.78) (4.24) (7.10) (5.33) (1.09) (2.35) (0.91) (2.40) (4.91) (3.61) Volume 1.44% 0.51% 0.43% 0.32% 0.14% -1.30% 0.80% 0.45% 0.31% 0.28% 0.04% -0.76% (6.03) (3.02) (3.15) (2.87) (1.09) (-4.48) (3.40) (2.49) (2.45) (2.29) (0.28) (-2.68) ZP 0.21% 0.29% 0.53% 0.83% 1.17% 0.96% 0.11% 0.34% 0.39% 0.45% 0.60% 0.49% (1.90) (2.65) (3.32) (4.80) (5.98) (3.94) (0.71) (2.43) (1.84) (2.05) (2.62) (1.72) CS -0.06% 0.19% 0.54% 0.64% 1.84% 1.90% -0.07% 0.18% 0.32% 0.52% 1.67% 1.74% (-0.75) (1.98) (4.43) (4.42) (8.03) (8.10) (-0.65) (1.22) (1.80) (2.32) (4.82) (4.76) C -0.03% 0.16% -0.05% -0.18% 1.19% 1.23% -0.02% 0.21% -0.10% -0.24% 0.77% 0.79% (-0.36) (1.25) (-0.32) (-1.28) (5.15) (4.87) (-0.14) (1.20) (-0.56) (-1.20) (2.52) (2.44) the lowest share prices, largest price impact (Amihud and C), lowest share turnover

and trading volume, highest bid-ask spreads (BIDASK and CS), and most zero return

trading days (ZP). For example, on an equal-weighted basis, the FRIC premium is

1.55% (t = 7.12) per month in the lowest price quintile while it is -0.08% (t = −0.80) in the highest price quintile, a difference of -1.63% per month that is highly significant

(t = −6.89). In addition, it is important to note that although the FRIC premium is more prominent among firms that are classified as illiquid based on measures used in the prior literature, it does not reside exclusively among those firms. For example, the FRIC premium is also significant among firms in the second and third lowest price quintiles (0.45% with a t-stat of 3.28 and 0.27% with a t-stat of 2.21, respec- tively). Likewise, the FRIC premium is 1.17% (t = 5.98) in the highest ZP quintile but is still significant at 0.83% (t = 4.80), 0.53% (t = 3.32), and 0.29% (t = 2.65) in the second, third, and fourth highest ZP quintiles, respectively. Results are sim- ilar using value-weighted returns. Thus, our evidence suggests that although FRIC is correlated with existing liquidity proxies, its return predictability is not merely a manifestation of previously documented liquidity effects as FRIC appears to con- tain additional information above and beyond those conventional liquidity proxies in predicting returns.

3.8 Anomalies and Frictions

We have argued that FRIC is a new measure of microstructure frictions that is both easy to construct and theoretically motivated. In this section, we study how it affects the inferences on asset pricing anomalies using two well-known examples, momentum (Jegadeesh and Titman(1993, 2001)) and idiosyncratic volatility (Ang,

125 Hodrick, Xing, and Zhang(2006)).

Empirical studies of the momentum and idiosyncratic volatility anomalies typi-

cally screen out firms with low prices to mitigate the impact of microstructure frictions

(see, e.g., Jegadeesh and Titman(2001) and Bali and Cakici(2008)), using price level

as a crude proxy for frictions. The choice of price filters is motivated by ease of use

combined with a lack of sufficiently long time-series data on more direct proxies of

microstructure frictions. As mentioned before, our FRIC measure is very easy to

construct as it only requires daily stock returns, and is readily available going back

to the 1920s. We next demonstrate the differentiated impact FRIC can have as a

measure of microstructure frictions by showing how the momentum and idiosyncratic

volatility anomalies are affected by screening out firms with high FRIC instead of low

prices.

For each anomaly, we eliminate the same number of firm/month observations

based on sorts by FRIC as we do when imposing the respective price filters (Price

< $1 and Price < $5). Panel A of Table 3.5 reports equal- and value-weighted mo- mentum decile returns under the different filters (FRIC and price). The average equal-weighted 10-1 momentum premium is always smaller when we use a FRIC fil- ter than when we use the corresponding price filter, and while the premium is still significantly positive under the FRIC filter it has a smaller t-statistic. For example,

the equal-weighted momentum premium is 1.31% (t = 5.40) per month under the $1

price filter but considerably lower at 0.89% (t = 3.27) per month under the corre-

sponding FRIC filter. We also report the q-factor alphas in the last column and note

that while the momentum premium under both $1 and $5 price filters produce signif-

icant positive alphas, the alphas under the alternative FRIC filters are much smaller

126 Table 3.5: Momentum, IVOL, and FRIC Equal- and value-weighted average returns (% per month) and their t-statistics (in parentheses) for (1) Momentum and IVOL sorted portfolios after screening stocks based on Price Level/FRIC, and (2) double-sorted portfolios on Momentum/IV and Price Level/FRIC are reported for the period July, 1966, to June, 2013. In Panel A, equal- and value-weighted average returns are reported for decile portfolios sorted on Momentum. At the end of each month, stocks are ranked by Momentum (cumulative average return over the past year skipping the most recent month) and sorted into deciles after screening out stocks based on Price Level (previous month’s stock price), or alternatively, FRIC. Price

127 screens drop stocks with previous month’s stock prices below $1 or $5, and corresponding FRIC screens drop the same number of stock/month observations of high-FRIC stocks. The difference in average returns between stocks in the highest and lowest Momentum deciles, and Hou, Xue, and Zhang(2015), in short HXZ, alphas of the 10-1 return spreads are reported as well. Panel B repeats Panel A for IVOL (standard deviation of residuals from Fama and French(1993) three-factor model using daily returns from previous month) sorted portfolios. In Panel C, equal- and value-weighted average returns are reported for double-sorted portfolios first sorted into quintiles on Price Level or FRIC, then on Momentum. At the end of each month, stocks are ranked by Price Level or FRIC and sorted into quintiles. Within each price/friction quintile, stocks are then sorted into quintiles based on their Momentum. The equal- and value-weighted average monthly return differences (% per month) between Momentum quintiles 5 and 1, or Momentum premiums, their HXZ alphas, and their t-statistics (in parentheses) are reported within each price/friction quintile. Then, the Momentum premium (and corresponding HXZ alpha) spreads between the high and low (low and high) FRIC (Price Level) quintiles and their associated t-statistics are also reported. Panel D repeats Panel C for IVOL instead of Momentum. Table 3.5 - Continued. Panel A: Momentum Portfolio Returns with Price vs. FRIC Screens Decile Portfolio 1 2 3 4 5 6 7 8 9 10 (10-1) HXZ Alpha Equal-weighted Price ≥ $1 0.41 0.84 0.98 1.04 1.12 1.19 1.30 1.42 1.61 1.72 1.31 0.63 (1.16) (2.95) (4.03) (4.56) (5.15) (5.59) (5.88) (6.14) (6.09) (5.25) (5.40) (2.62) High FRIC Screen 0.83 0.85 0.99 1.05 1.09 1.17 1.30 1.42 1.59 1.72 0.89 0.05 (2.15) (2.93) (3.98) (4.58) (5.05) (5.50) (5.92) (6.13) (6.07) (5.23) (3.27) (0.20)

Price ≥ $5 0.27 0.82 0.96 1.00 1.09 1.17 1.25 1.38 1.56 1.74 1.47 0.98 (0.90) (3.31) (4.28) (4.78) (5.35) (5.69) (5.91) (6.12) (6.07) (5.34) (6.26) (4.11)

128 High FRIC Screen 0.60 0.86 0.97 1.00 1.05 1.16 1.26 1.38 1.52 1.65 1.05 0.32 (1.62) (3.09) (4.06) (4.52) (5.03) (5.54) (5.79) (6.05) (5.87) (4.98) (3.92) (1.20)

Value-weighted Price ≥ $1 0.21 0.55 0.96 0.76 0.84 0.83 1.12 1.07 1.21 1.51 1.30 0.51 (0.59) (1.91) (4.00) (3.67) (4.28) (4.23) (5.50) (5.00) (4.93) (4.76) (4.14) (1.64) High FRIC Screen 0.25 0.55 0.90 0.78 0.84 0.82 1.09 1.10 1.20 1.49 1.24 0.39 (0.66) (1.89) (3.69) (3.67) (4.27) (4.21) (5.37) (5.15) (4.90) (4.72) (3.76) (1.22)

Price ≥ $5 0.30 0.76 0.86 0.80 0.83 0.88 1.05 1.08 1.19 1.53 1.23 0.64 (0.95) (2.97) (4.08) (4.16) (4.36) (4.48) (5.18) (5.04) (4.86) (4.79) (4.18) (2.15) High FRIC Screen 0.14 0.66 0.99 0.75 0.90 0.81 1.04 1.11 1.18 1.52 1.38 0.66 (0.37) (2.34) (4.37) (3.76) (4.63) (4.14) (5.15) (5.17) (4.84) (4.74) (4.16) (1.99)

Continued. Table 3.5 - Continued. Panel B: IVOL Portfolio Returns with Price vs. FRIC Screens Decile Portfolio 1 2 3 4 5 6 7 8 9 10 (10-1) HXZ Alpha Equal-weighted Price ≥ $1 1.06 1.21 1.28 1.35 1.40 1.31 1.33 1.19 0.97 0.36 -0.70 -0.25 (7.32) (6.63) (6.28) (6.16) (5.85) (4.98) (4.61) (3.81) (2.89) (0.99) (-2.42) (-1.37) High FRIC Screen 1.12 1.19 1.28 1.34 1.40 1.31 1.31 1.26 1.02 0.65 -0.47 0.18 (7.58) (6.58) (6.35) (6.11) (5.86) (5.03) (4.57) (3.99) (2.98) (1.62) (-1.47) (0.86)

Price ≥ $5 1.06 1.21 1.25 1.32 1.33 1.39 1.21 1.21 0.96 0.16 -0.90 -0.63 (7.47) (6.84) (6.42) (6.24) (5.89) (5.75) (4.61) (4.29) (3.11) (0.49) (-3.67) (-4.51)

129 High FRIC Screen 1.09 1.17 1.26 1.29 1.30 1.37 1.20 1.21 0.99 0.39 -0.71 -0.06 (7.64) (6.70) (6.45) (6.05) (5.65) (5.45) (4.33) (3.92) (2.91) (0.97) (-2.19) (-0.28)

Value-weighted Price ≥ $1 0.93 0.99 0.96 1.01 0.90 1.00 0.82 0.70 0.22 -0.23 -1.17 -0.65 (5.97) (5.48) (4.62) (4.46) (3.57) (3.59) (2.74) (2.07) (0.57) (-0.58) (-3.38) (-2.80) High FRIC Screen 0.95 0.98 0.98 0.98 0.90 1.01 0.89 0.68 0.21 -0.03 -0.97 -0.35 (6.01) (5.50) (4.73) (4.32) (3.62) (3.58) (2.96) (2.00) (0.58) (-0.06) (-2.54) (-1.30)

Price ≥ $5 0.93 1.00 0.94 0.96 0.98 0.91 0.91 0.79 0.61 -0.28 -1.21 -0.81 (5.91) (5.71) (4.82) (4.47) (4.23) (3.63) (3.27) (2.63) (1.83) (-0.78) (-4.15) (-4.25) High FRIC Screen 0.94 0.99 0.99 0.96 0.96 0.97 0.89 0.71 0.37 -0.06 -0.99 -0.40 (5.95) (5.66) (5.07) (4.49) (4.07) (3.84) (3.09) (2.26) (1.02) (-0.13) (-2.65) (-1.46)

Continued. Table 3.5 - Continued. Panel C: Momentum Premiums Across Price vs. FRIC Quintiles Momentum Low 2 3 4 High High- HXZ Low 2 3 4 High High- HXZ Quintiles Low Alpha Low Alpha Equal-weighted Value-weighted Low 2.34 1.75 1.83 1.89 1.90 -0.44 -1.38 1.29 1.04 1.32 1.42 1.35 0.07 -0.81 (4.37) (4.51) (4.97) (5.47) (5.02) (-1.45) (-4.57) (1.94) (2.14) (3.08) (3.58) (3.14) (0.15) (-1.68) 2 0.15 0.87 1.17 1.36 1.63 1.48 0.92 0.45 1.06 1.03 1.34 1.46 1.01 0.43 (0.43) (3.05) (4.43) (5.03) (4.90) (6.95) (4.21) (1.06) (3.01) (3.30) (4.06) (3.96) (3.44) (1.42) Price Quintiles 3 0.55 0.98 1.17 1.37 1.63 1.08 0.60 0.59 0.94 1.13 1.42 1.48 0.90 0.27 (1.92) (4.13) (5.30) (5.82) (5.55) (5.75) (3.16) (1.78) (3.45) (4.40) (5.43) (4.77) (3.83) (1.14) 4 0.71 1.00 1.19 1.33 1.67 0.97 0.56 0.69 0.94 1.08 1.15 1.51 0.81 0.11 (2.82) (4.93) (6.01) (6.18) (5.88) (5.14) (3.04) (2.46) (4.35) (5.09) (5.13) (5.28) (3.57) (0.49) High 0.73 0.86 1.01 1.13 1.57 0.84 0.70 0.81 0.74 0.87 1.01 1.31 0.50 0.22 (3.44) (4.64) (5.44) (5.41) (5.43) (3.85) (3.35) (3.78) (4.12) (4.59) (4.83) (4.77) (2.09) (0.95) 130 Low-High -1.28 -2.09 -0.43 -1.04 (-4.10) (-7.19) (-0.93) (-2.20)

High 2.06 1.34 1.54 1.74 1.97 -0.09 -1.05 1.15 0.63 1.00 1.16 1.33 0.18 -0.70 (4.44) (4.12) (5.32) (6.05) (5.97) (-0.32) (-3.67) (2.41) (1.83) (3.10) (4.27) (4.13) (0.46) (-1.71) 4 1.10 1.18 1.20 1.40 1.73 0.63 -0.10 0.89 0.96 0.91 0.94 1.27 0.38 -0.30 (3.32) (5.13) (5.76) (6.59) (6.26) (2.72) (-0.44) (2.70) (3.79) (4.26) (4.33) (4.70) (1.38) (-1.08) FRIC Quintiles 3 1.00 1.08 1.14 1.29 1.70 0.70 0.14 0.82 0.82 0.74 1.03 1.42 0.60 0.05 (3.44) (5.21) (6.01) (6.66) (6.94) (3.52) (0.72) (3.00) (3.92) (3.98) (5.28) (6.03) (2.47) (0.19) 2 0.69 0.98 1.11 1.30 1.72 1.02 0.45 0.64 0.88 0.82 1.15 1.37 0.73 0.10 (2.43) (4.44) (5.46) (6.15) (6.69) (5.48) (2.46) (2.22) (4.27) (4.12) (5.48) (5.42) (2.97) (0.42) Low 0.34 0.71 0.97 1.24 1.44 1.10 0.52 0.01 0.59 0.91 0.92 1.29 1.28 0.68 (0.88) (2.39) (3.46) (4.22) (4.04) (4.34) (2.04) (0.02) (2.06) (3.46) (3.34) (3.70) (3.83) (2.01) High-Low -1.19 -1.57 -1.10 -1.38 (-5.18) (-6.48) (-2.81) (-3.35)

Continued. Table 3.5 - Continued. Panel D: IVOL Discounts Across Price vs. FRIC Quintiles IVOL Low 2 3 4 High High- HXZ Low 2 3 4 High High- HXZ Quintiles Low Alpha Low Alpha Equal-weighted Value-weighted Low 1.84 1.94 2.03 1.93 1.97 0.14 0.47 1.19 1.79 1.32 1.37 0.66 -0.53 0.11 (6.03) (5.32) (5.04) (4.44) (3.96) (0.49) (1.76) (3.10) (3.90) (2.56) (2.33) (1.03) (-1.19) (0.25) 2 1.33 1.60 1.33 0.89 -0.06 -1.39 -1.09 1.44 1.54 1.11 0.59 -0.14 -1.58 -1.07 (5.98) (5.75) (4.27) (2.73) (-0.17) (-7.32) (-7.10) (4.99) (4.57) (2.92) (1.40) (-0.29) (-5.07) (-3.65) Price Quintiles 3 1.24 1.43 1.41 1.16 0.40 -0.84 -0.67 1.21 1.35 1.50 0.98 -0.03 -1.23 -1.00 (6.84) (6.23) (5.53) (4.07) (1.28) (-4.28) (-4.79) (5.61) (5.16) (4.95) (2.94) (-0.08) (-4.90) (-4.85) 4 1.18 1.37 1.34 1.24 0.74 -0.44 -0.29 1.13 1.11 1.06 0.91 0.30 -0.83 -0.49 (7.30) (6.69) (5.96) (4.97) (2.48) (-2.20) (-2.24) (6.42) (4.95) (4.33) (3.20) (0.91) (-3.36) (-2.75) High 0.99 1.10 1.12 1.06 1.00 0.01 0.31 0.87 0.97 0.93 0.84 0.74 -0.13 0.26 131 (6.53) (6.01) (5.51) (4.58) (3.40) (0.04) (1.99) (5.54) (5.31) (4.43) (3.63) (2.57) (-0.60) (1.69) Low-High 0.13 0.16 -0.40 -0.15 (0.47) (0.56) (-0.94) (-0.32)

High 1.48 1.71 1.76 1.66 2.07 0.59 1.23 1.04 1.20 1.18 0.76 1.15 0.10 0.57 (6.58) (5.95) (5.27) (4.46) (4.43) (1.85) (4.32) (4.02) (3.49) (3.05) (1.74) (2.21) (0.24) (1.37) 4 1.19 1.40 1.49 1.44 1.08 -0.10 0.26 0.95 1.10 1.23 0.99 0.44 -0.51 -0.31 (7.07) (6.86) (6.48) (5.11) (3.09) (-0.42) (1.57) (4.98) (4.71) (4.53) (3.19) (1.16) (-1.77) (-1.32) FRIC Quintiles 3 1.18 1.23 1.34 1.34 1.11 -0.07 0.16 0.96 0.93 0.99 1.00 0.57 -0.39 -0.12 (7.78) (6.64) (6.50) (5.47) (3.47) (-0.32) (1.19) (5.83) (4.84) (4.50) (3.84) (1.71) (-1.49) (-0.60) 2 1.07 1.25 1.31 1.25 0.91 -0.16 0.15 0.90 1.04 1.01 0.88 0.59 -0.31 -0.07 (6.68) (6.20) (5.73) (4.82) (2.88) (-0.73) (1.21) (5.26) (5.18) (4.42) (3.29) (1.75) (-1.22) (-0.36) Low 1.15 1.14 1.07 0.91 0.23 -0.92 -0.06 0.97 0.89 0.60 0.31 -0.49 -1.46 -0.80 (5.38) (4.16) (3.34) (2.51) (0.55) (-3.12) (-0.27) (4.47) (3.12) (1.81) (0.77) (-1.07) (-4.05) (-2.73) High-Low 1.51 1.29 1.56 1.37 (6.64) (5.28) (3.67) (2.99) in magnitude and statistically insignificant. The differences in momentum premium

and q-factor alpha between the different filters are less pronounced for value-weighted returns.

Similarly, Panel B of Table 3.5 reports the idiosyncratic volatility (IVOL) decile returns and q-factor alphas under the different filters (FRIC and price). The 10-1

IVOL discount is uniformly smaller in magnitude under the FRIC filters than un- der the corresponding price filters, regardless of the weighting method. Moreover, while the q-factor alpha is statistically significant in three out of four cases under

the price filters, it is always insignificant under the FRIC filters. Hence, in line with

Hou, Xue, and Zhang(2015), we find that the idiosyncratic volatility anomaly can

be completely explained by the q-factor model when we screen out stocks that suffer

severe microstructure frictions.

In Panels C and D of Table 3.5, we further examine how the momentum and

idiosyncratic volatility anomalies are affected by microstructure frictions by conduct-

ing double-sorts instead of pre-screening the data. Specifically, the two panels report

equal- and value-weighted average returns and q-factor alphas for double-sorted port-

folios first on price (alternatively, FRIC) and then on momentum or IVOL. The

message from Panel C is that there is no significant momentum premium in either

the lowest price quintile or the highest FRIC quintile regardless of how returns are

weighted. More importantly, the effect of FRIC on the momentum premium is gen-

erally monotonic but such is not the case for price. For example, the equal-weighted

5-1 momentum premium increases monotonically from -0.09% (t = −0.32) in the

highest FRIC quintile to 1.10% (t = 4.34) in the lowest FRIC quintile. By contrast,

the momentum premium first increases from -0.44% (t = −1.45) in the lowest price

132 quintile to 1.48% (t = 6.95) in the second lowest price quintile and then gradually

declines to 0.84% (t = 3.85) in the highest price quintile. We observe similar patterns

in q-factor alphas across the FRIC and price quintiles.

Panel D studies the idiosyncratic volatility anomaly under the alternative price versus FRIC sorts. First, there is no significant IVOL discount in either the lowest price quintile or the highest FRIC quintile regardless of the weighting scheme. Second, the IVOL discount is related to price in an inverted U-shaped pattern with signifi- cant discounts in quintiles two, three, and four for both equal- and value-weighted returns. By contrast, the IVOL discount is concentrated in the lowest FRIC quintile, regardless of weighting scheme. Again, similar patterns are observed for the q-factor

alphas.

Overall, our results in this section suggest that screening/sorting based on our

FRIC measure provides cleaner inferences on asset pricing anomalies such as the

momentum and idiosyncratic volatility effects. These results suggest that standard

price level filters commonly used in the literature represent a poor way of excluding

extreme friction firms. Furthermore, the results also suggest that we may ultimately

pin down the sources of the two anomalies by properly identifying and focusing on

firms that are not affected by severe microstructure frictions.

3.9 Extension: Results for the United Kingdom

One of the main advantages of our FRIC measure is that it is very simple to

construct, and therefore can be applied in many settings where data on trading vol-

ume, bid-ask prices, and daily high-low prices are either unavailable or unreliable. To

demonstrate this, we study the effect of microstructure frictions on the cross section of

133 stock returns in the United Kingdom (UK) using Thomson’s Datastream/Worldscope data from July, 1981 to June, 2013.36 This analysis also serves as an out-of-sample test for the results thus far obtained using the US CRSP/Compustat sample.

The UK market is a particularly interesting setting because there are a number of known issues regarding the availability and quality of data required to compute various microstructure liquidity measures. For example, volume data are available for a reasonable number of firms only from the late 1980s onward, while bid-ask prices and intraday high-low prices are only available after the end of 1986. Such limitations make it difficult to compute liquidity proxies such as the Amihud(2002) illiquidity measure, bid-ask spreads, and Corwin and Schultz(2012) measure over an extended time period. Furthermore, Lee(2011) shows that using the zero return proportion measure, UK firms appear more illiquid (have more zero-return trading days) not only compared to other developed markets but even than most emerging markets.

Lee(2011) also shows that zero return proportion is a negative cross-sectional return predicator for UK firms.37 That is, firms with more zero return trading days have lower average returns, which is counter-intuitive. We investigate whether our FRIC measure can resolve these issues.

Panel A of Table 3.6 reports summary statistics for our FRIC measure along with four commonly used microstructure liquidity proxies: Zero return proportion (ZP),

Amihud(2002) illiquidity measure (Amihud), Bid-Ask Spread (BIDASK), and Cor- win and Schultz(2012) measure (CS). Consistent with Lee(2011), we also find that

36We obtain data from Datasteam/Worldscope and apply several screening procedures for security names to remove non-common shares, and for monthly and daily returns, as suggested by Ince and Porter(2006), Griffin, Kelly, and Nardari(2010), and Lee(2011). Also following Lee(2011), we drop firm-month observations where ZP is more than 80%. We also follow Hou, Karolyi, and Kho (2011) to construct size, book-to-market equity, cash flow-to-price ratio, and momentum. 37See Table 3 of Lee(2011)

134 Table 3.6: Illiquidity and the Cross-Section of Stock Returns in the United Kingdom

This table reports summary statistics of different liquidity proxies and results from Fama and Mac- Beth(1973) regressions using stocks listed in the United Kingdom. Panel A reports the time-series averages of cross-sectional statistics for the friction measure, FRIC, and other liquidity proxies that are commonly used in the literature. In addition, Panel B reports the time-series averages of cross- sectional correlations between different liquidity proxies. Finally, Panel C reports the time-series average coefficients and their t-statistics (in parentheses) from monthly Fama and MacBeth(1973) cross-sectional regressions of individual stock returns on various firm-level characteristics and liq- uidity variables. Firm-level variables that are used in this analysis are log of firm size (market capitalization), log of book-to-market equity (BE/ME), CFP+, CFP/- dummy (if cash flow is posi- tive, CFP+ equals cash flow over price and the CFP/- dummy is 0. If cash flow is not positive, CFP+ is 0 and the CFP/- dummy is 1), momentum (cumulative raw return from month t − 6 to month t − 2), log of prior month’s share price, prior month’s return (Ret−1:−1), FRIC (constructed using previous 1-year daily return data), zero return proportion measure (ZP: number of days with zero return divided by number of trading days from previous year), Amihud(2002) illiquidity measure (daily absolute return divided by daily dollar trading volume, averaged over prior year), average daily effective bid-ask spread, and Corwin and Schultz(2012) spread estimate (CS). The sample covers the period from July, 1981 to June, 2013. 1: From July, 1990 to June, 2013 2: From July, 1988 to June, 2013 3: From July, 1988 to June, 2013

135 Table 3.6 - Continued. Panel A: Comparison between liquidity proxies Mean Standard 25th 50th 75th deviation percentile percentile percentile FRIC (basis points) -0.36 1.39 -0.86 -0.31 0.15 ZP 0.37 0.22 0.19 0.38 0.53 Amihud 9.12 134.90 0.00 0.04 0.18 BIDASK (%) 4.33 4.26 1.36 3.04 5.75 CS (%) 1.40 1.76 0.50 0.87 1.61

Panel B: Correlations between liquidity proxies FRIC ZP Amihud BIDASK CS FRIC 1.00 ZP -0.04 1.00 Amihud 0.03 0.10 1.00 BIDASK -0.10 0.58 0.24 1.00 CS 0.01 0.37 0.17 0.64 1.00

Panel C: Fama-MacBeth Regressions (1) (2) (3) (4) (5) (6) (7) Log(Size) -0.03 -0.03 -0.06 0.03 0.02 0.01 -0.05 (-0.60) (-0.63) (-1.40) (0.60) (0.51) (0.32) (-0.86) Log(BE/ME) 0.17 0.17 0.17 0.17 0.12 0.15 0.16 (3.21) (3.18) (3.21) (2.62) (2.11) (2.57) (2.66) CFP+ 1.41 1.40 1.39 1.66 1.80 1.75 1.55 (5.06) (5.04) (5.01) (4.32) (5.00) (5.13) (4.08) CFP/- dummy -0.34 -0.33 -0.32 -0.22 -0.32 -0.33 -0.34 (-1.98) (-1.93) (-1.89) (-1.10) (-1.96) (-2.11) (-1.55) Momentum 1.54 1.60 1.59 1.40 1.47 1.47 1.40 (6.93) (7.14) (7.14) (5.38) (5.92) (6.35) (5.42) Log(Price) 0.01 0.01 0.02 0.09 0.07 0.08 0.08 (0.08) (0.14) (0.27) (1.36) (0.97) (1.13) (1.10) Ret−1:−1 -1.34 -1.34 -1.29 -0.65 -0.52 -0.38 -0.72 (-2.93) (-2.94) (-2.87) (-1.12) (-1.13) (-0.84) (-1.23) FRIC 9.63 10.06 (2.63) (2.39) ZP -0.10 -0.49 (-0.23) (-1.04) Amihud1 -0.20 -1.77 (-0.19) (-0.95) BIDASK2 -0.97 -1.79 (-0.45) (-0.70) CS3 -5.73 -2.17 (-1.45) (-0.34)

136 UK firms on average have high proportions of zero return trading days. Specifically,

UK firms in our sample have on average 37% zero return trading days in a given year, while US firms have 18% zero return trading days on average. On the other hand, the average FRIC for UK firms is -0.36 basis points, which is not much different from the average FRIC of US firms, 1.9 basis points. Therefore, while the ZP measure suggests that the UK market is much more illiquid than the US market, the FRIC measure actually implies that there is little difference in microstructure frictions between the two markets.

Given that our FRIC measure and other liquidity proxies seem to paint different pictures for UK firms, it is of interest to examine their correlations, which are shown in Panel B of Table 3.6. The four conventional liquidity proxies are positively corre- lated with each other, although the Amihud measure, which is positively skewed to a much greater extent compared to other measures, is only weakly correlated with ZP or CS. On the other hand, FRIC is very weakly correlated with conventional liquidity proxies, indicating that its effect on UK stock returns will likely be different from that of other measures.

Panel C of Table 3.6 examines the return predictability of FRIC and other liquidity proxies using Fama-MacBeth regressions. Results in the first column are consistent with prior studies in that average returns are negatively related to size and previous month’s returns, and positively related to BE/ME, Cash flow-to-Price ratio, and Mo- mentum (previous six-month’s returns).38 Columns (2) through (6) add FRIC and different liquidity proxies one at a time to those standard control variables, and the

38See Hou, Karolyi, and Kho(2011) for a comprehensive review of these findings in an international setting.

137 last column includes all variables together. To ensure there are enough firms to esti- mate each cross-sectional regression, we begin the sample in July, 1990 for regressions including Amihud, whereas regressions including BIDASK or CS begin in July, 1988.

Similar to the evidence from US firms, column (2) confirms that UK firms with larger

FRIC tend to have higher average returns after controlling for standard return pre- dictors. The coefficient on FRIC is 9.63 (t-statistic = 2.63). In untabulated results, we find that the average FRIC spread in the UK is 4.58 basis points between extreme

FRIC-sorted decile portfolios. Hence, the coefficient on FRIC implies a sizable dif- ference in returns between the extreme deciles of 0.44% per month.

On the other hand, columns (3) to (6) show inconsistent results regarding the liq- uidity premiums associated with traditional proxies. We find statistically insignificant effects for ZP, Amihud, BIDASK, and CS, where the signs are all negative instead of positive. For example, the coefficient on ZP is -0.10 (t-statistic = −0.23), while it is

-0.20 (t-statistic = −0.19) for Amihud. These negative coefficients are inconsistent with the idea that investors demand a premium for holding illiquid firms.

When we include all variables in column (7), FRIC is still the only liquidity-related variable that sustains a large and significantly positive coefficient. Coefficients on ZP,

Amihud, BIDASK, and CS remain negative and statistically insignificant. Overall, the regression results suggest that among the five liquidity-related variables (FRIC,

ZP, Amihud, BIDASK, and CS), our FRIC measure is the only proxy that reliably uncovers a consistent premium for UK firms that is both economically large and statistically significant.

138 3.10 Conclusion

In this paper, we propose a parsimonious measure to characterize the severity of microstructure frictions at the individual stock level and assess the impact of this measure on the cross section of stock returns. Our friction measure is directly mo- tivated by the key insight from Blume and Stambaugh(1983) that microstructure frictions can introduce an upward bias in average observed stock returns. Our mea- sure, which we call FRIC, essentially captures the degree of this upward bias at the individual stock level.

Our FRIC measure is very easy to construct, requiring only daily stock return data as input. It is based on the premise that one can obtain a “friction-free” esti- mate of the true average daily return by dividing the average two-day observed return by the average lagged one-day observed return, which effectively cancels out the mi- crostructure friction-induced common price noise that infects both the numerator and denominator. We thus define FRIC as the difference between the observed average daily return and the friction-free average daily return.

Using our FRIC measure, we find that firms with large microstructure frictions are low priced, small, and volatile, consistent with economic priors. In addition, we find that FRIC is strongly correlated with commonly used microstructure liquid- ity proxies such as Amihud, CS, ZP, dollar volume, share turnover, bid-ask spread, and the Hasbrouck(2009) C-measure. We also examine the time-series pattern of the market-wide FRIC measure and find that it displays considerable variation over time. For instance, spikes in FRIC are clearly visible during the recessions of 1974 to

1975 and 1990 to 1991, the tech bubble of the late 1990s, and the financial crisis of

2007-2009.

139 On an equal- (value-) weighted basis, stocks in the highest FRIC decile command a return premium as large as 16.29% (10.34%) per year. This finding is consistent with the idea that microstructure frictions are both costly and risky to investors. Our results are robust to a variety of return adjustments for common risk factors, alter- native specifications, subsamples, and subperiod analysis. The friction premium also persists after controlling for standard variables known to explain stock returns and other conventional liquidity proxies in Fama and MacBeth(1973) regressions. We also find that the friction premium is more pronounced among small, value, volatile, past loser, and illiquid firms, which is consistent with the consensus that microstruc- ture frictions present a more serious problem among these firms.

We use two well-known anomalies, momentum and idiosyncratic volatility, to demonstrate the difference between our FRIC measure and other conventional meth- ods of controlling for the effects of microstructure frictions on asset pricing inferences, such as price level filters. We find that screening out firms with high FRIC produces uniformly smaller and less significant momentum and idiosyncratic volatility anoma- lies compared to results based on price filters that are commonly used in the literature.

Furthermore, momentum premiums increase monotonically from high to low FRIC quintiles and the idiosyncratic volatility discount is primarily contained in the lowest

FRIC quintile, whereas the patterns for the two anomalies are inverted U-shaped across price quintiles. Together, these results suggest that price filters as a way of handling extreme friction firms may be misguided, and that FRIC is a more reliable measure in this regard.

Finally, our FRIC measure, unlike other conventional liquidity proxies, consis- tently produces a positive return premium that is both economically and statistically

140 significant for the cross section of UK stock returns, highlighting its usefulness in set- tings where the availability or quality of data on volume, bid-ask prices, and intraday high-low prices may be suspect.

Overall, our paper provides a theoretically-motivated, yet simple and practical method of capturing the severity of microstructure frictions in individual stocks, which we hope will be useful for research in market microstructure, market frictions, and asset pricing in general.

141 Bibliography

Acharya, Viral V., Heitor Almeida, and Murillo Campello, 2007. Is cash negative debt? A hedging perspective on corporate financial policies. Journal of Financial Intermediation, Vol. 16, No. 4, pp. 515-554. Acharya, Viral V. and Lasse Heje Pedersen, 2005. Asset pricing with liquidity risk. Journal of Financial Economics, Vol. 77, pp. 375-410. Almeida, Heitor, Murillo Campello, and Michael S. Weisbach, 2004. The cash flow sensitivity of cash. The Journal of Finance, Vol. 59, No. 4, pp. 1777-1804. Amihud, Yakov, 2002. Illiquidity and stock returns: Cross-section and time-series effects. Journal of Financial Markets, Vol. 5, pp. 31-56. Amihud, Yakov and Haim Mendelson, 1986. Asset pricing and the bid-ask spread. Journal of Financial Economics, Vol. 17, pp. 223-249. Ang, Andrew, Robert J. Hodrick, Yuhang Xing, and Xiaoyan Zhang, 2006. The cross- section of volatility and expected returns. The Journal of Finance, Vol. 61, No. 1, pp. 259-299. Angel, James J., Lawrence E. Harris, and Chester S. Spatt, 2011. Equity trading in the 21st century. The Quarterly Journal of Finance, Vol. 1, No. 1, pp. 1-53. Angel, James J., Lawrence E. Harris, and Chester S. Spatt, 2015. Equity trading in the 21st century: An update. The Quarterly Journal of Finance, Vol. 5, No. 1, pp. 1-39. Asparouhova, Elena, Hendrik Bessembinder, and Ivalina Kalcheva, 2010. Liquidity biases in asset pricing tests. Journal of Financial Economics, Vol. 96, pp. 215- 237. Asparouhova, Elena, Hendrik Bessembinder, and Ivalina Kalcheva, 2013. Noisy prices and inference regarding returns. The Journal of Finance, Vol. 68, No. 2, pp. 665- 714. Azar, Jos´e,Martin C. Schmalz, and Isabel Tecu, 2016. Anti-competitive effects of common ownership. Ross School of Business Working Paper, No. 1235.

142 Bali, Turan G. and Nusret Cakici, 2008. Idiosyncratic volatility and the cross-section of expected returns. Journal of Financial and Quantitative Analysis, Vol. 43, No. 1, pp. 29-58.

Bates, Thomas W., Kathleen M. Kahle, and Ren´eM. Stulz, 2009. Why do U.S. firms hold so much more cash than they used to? The Journal of Finance, Vol. 64, No. 5, pp. 1985-2021.

Bebchuk, Lucian A., Alon Brav, and Wei Jiang, 2015. The long-term effects of hedge fund activism. Columbia Law Review, Vol. 115, No. 5, pp. 1085-1156.

Bekaert, Geert, Campbell R. Harvey, and Christian Lundblad, 2007. Liqudity and expected returns: Lessons from emerging markets. Review of Financial Studies, Vol. 20, No. 5, pp. 1783-1831.

Berger, Philip G. and Eli Ofek, 1995. Diversification’s effect on firm value. Journal of Financial Economics, Vol. 37, pp. 39-65.

Bernheim, B. Douglas and Michael D. Whinston, 1990. Multimarket contact and collusive behavior. The RAND Journal of Economics, Vol. 21, No. 1, pp. 1-26.

Berry, Steven T., 1992. Estimation of a model of entry in the airline industry. Econo- metrica, Vol. 60, No. 4, pp. 889-917.

Bethel, Jennifer E., Julia Porter Liebeskind, and Tim Opler, 1998. Block share pur- chases and corporate performance. The Journal of Finance, Vol. 53, No. 2, pp. 605-634.

Billett, Matthew T. and David C. Mauer, 2003. Cross-subsidies, external financing constraints, and the contribution of the internal capital market to firm value. Review of Financial Studies, Vol. 16, No. 4, pp. 1167-1201.

Blume, Marshall E. and Robert F. Stambaugh, 1983. Biases in computed returns: An application to the size effect. Journal of Financial Economics, Vol. 12, pp. 387-404.

Bolton, Patrick and David S. Scharfstein, 1990. A theory of predation based on agency problems in financial contracting. The American Economic Review, Vol. 80, No. 1, pp. 93-106.

Brander, James A. and Tracy R. Lewis, 1986. Oligopoly and financial structure: The limited liability effect. The American Economic Review, Vol. 76, No. 5, pp. 956- 970.

Brav, Alon, Wei Jiang, and Hyunseob Kim, 2015. The real effects of hedge fund activism: Productivity, asset allocation, and labor outcomes. Review of Financial Studies, Vol. 28, No. 10, pp. 2723-2769.

143 Brav, Alon, Wei Jiang, Frank Partnoy, and Randall Thomas, 2008. Hedge fund ac- tivism, corporate governance, and firm performance. The Journal of Finance, Vol. 63, No. 4, pp. 1729-1775.

Brennan, Michael J., Tarun Chordia, and Avanidhar Subrahmanyam, 1998. Alter- native factor specifications, security characteristics, and the cross-section of ex- pected stock returns. Journal of Financial Economics, Vol. 49, pp. 345-373.

Brennan, Michael J. and Avanidhar Subrahmanyam, 1996. Market microstructure and asset pricing: On the compensation for illiquidity in stock returns. Journal of Financial Economics, Vol. 41, pp. 441-464.

Brennan, Michael J. and Ashley Wang, 2010. The mispricing return premium. Review of Financial Studies, Vol. 23, No. 9, pp. 3437-3468.

Bulow, Jeremy I., John D. Geanakoplos, and Paul D. Klemperer, 1985. Multimarket oligopoly: Strategic substitutes and complements. Journal of Political Economy, Vol. 93, No. 3, pp. 488-511.

Busse, Meghan, 2002. Firm financial condition and airline price wars. The RAND Journal of Economics, Vol. 33, No. 2, pp. 298-318.

Campa, Jose Manuel and Simi Kedia, 2002. Explaining the diversification discount. The Journal of Finance, Vol. 57, No. 4, pp. 1731-1762.

Campello, Murillo, 2003. Capital structure and product markets interactions: Evi- dence from business cycles. Journal of Financial Economics, Vol. 68, pp. 353-378.

Carhart, Mark M., 1997. On persistence in mutual fund performance. The Journal of Finance, Vol. 52, No. 1, pp. 57-82.

Chevalier, Judith A., 1995. Capital structure and product market competition: Em- pirical evidence from the supermarket industry. The American Economic Review, Vol. 85, No. 3, pp. 415-435.

Chevalier, Judith A., 1995. Do LBO supermarkets charge more? An empirical analysis of the effects of LBOs on supermarket pricing. The Journal of Finance, Vol. 50, No. 4, pp. 1095-1112.

Chevalier, Judith A., 2004. What do we know about cross-subsidization? Evidence from merging firms. The B.E. Journal of Economic Analysis and Policy, Vol. 4, No. 1, pp. 1-29.

Chevalier, Judith A. and David S. Scharfstein, 1996. Capital market imperfections and countercyclical markups: Theory and evidence. The American Economic Review, Vol. 86, No. 4, pp. 703-725.

144 Ciliberto, Federico and Elie Tamer, 2009. Market structure and multiple equilibria in airline markets. Econometrica, Vol. 77, No. 6, pp. 1791-1828.

Clifford, Christopher P., 2008. Value creation or destruction? Hedge funds as share- holder activists. Journal of Corporate Finance, Vol. 14, pp. 323-336.

Cohen, Randolph B., Paul A. Gompers, and Tuomo Vuolteenaho, 2002. Who un- derreacts to cash-flow news? Evidence from trading between individuals and institutions. Journal of Financial Economics, Vol. 66, pp. 409-462.

Comment, Robert and Gregg A. Jarrell, 1995. Corporate focus and stock returns. Journal of Financial Economics, Vol. 37, pp. 67-87.

Cookson, J. Anthony, 2017. Leverage and strategic preemption: Lessons from entry plans and incumbent investments. Journal of Financial Economics, Vol. 123, No. 2, pp. 292-312.

Cooper, Michael J., Huseyin Gulen, and Michael J. Schill, 2008. Asset growth and the cross-section of stock returns. The Journal of Finance, Vol. 63, No. 4, pp. 1609-1651.

Corwin, Shane A. and Paul Schultz, 2012. A simple way to estimate bid-ask spreads from daily high and low prices. The Journal of Finance, Vol. 67, No. 2, pp. 719-760.

Cust´odio,Cl´audia,2014. Mergers and acquisitions accounting and the diversification discount. The Journal of Finance, Vol. 69, No. 1, pp. 219-240.

DeAngelo, Harry and Linda DeAngelo, 1990, Dividend policy and financial distress: An empirical investigation of troubled NYSE firms. The Journal of Finance, Vol. 45, No. 5, pp. 1415-1431.

DeAngelo, Harry, Linda DeAngelo, and Douglas J. Skinner, 1992, Dividends and losses. The Journal of Finance, Vol. 47, No. 5, pp. 1837-1863.

Denis, David J. and Stephen B. McKeon, 2012. Debt financing and financial flexibility: Evidence from proactive leverage increases. Review of Financial Studies, Vol. 25, No. 6, pp. 1897-1929.

Denis, David J. and Valeriy Sibilkov, 2010. Financial constraints, investment, and the value of cash holdings. Review of Financial Studies, Vol. 23, No. 1, pp. 247-269.

Diether, Karl B., Kuan-Hui Lee, and Ingrid M. Werner, 2009. Short-sale strategies and return predictability. Review of Financial Studies, Vol. 22, No. 2, pp. 575- 607.

145 Duchin, Ran, Oguzhan Ozbas, and Berk A. Sensoy, 2010. Costly external finance, cor- porate investment, and the subprime mortgage credit crisis. Journal of Financial Economics, Vol. 97, pp. 418-435. Duchin, Ran and Denis Sosyura, 2013. Divisional managers and internal capital mar- kets. The Journal of Finance, Vol. 68, No. 2, pp. 387-429. Ertimur, Yonca, Fabrizio Ferri, and Volkan Muslu, 2011. Shareholder activism and CEO pay. Review of Financial Studies, Vol. 24, No. 2, pp. 535-592. Evans, William N. and Ioannis N. Kessides, 1994. Living by the “Golden rule”: Multi- market contact in the U.S. airline industry. The Quarterly Journal of Economics, Vol. 109, No. 2, pp. 341-366. Fama, Eugene F. and James D. MacBeth, 1973. Risk, return, and equilibrium: Em- pirical tests. Journal of Political Economy, Vol. 81, No. 3, pp. 607-636. Fama, Eugene F. and Kenneth R. French, 1992. The cross-section of expected stock returns. The Journal of Finance, Vol. 47, No. 2, pp. 427-465. Fama, Eugene F. and Kenneth R. French, 1993. Common risk factors in the returns on stocks and bonds. Journal of Financial Economics, Vol. 33, pp. 3-56. Fama, Eugene F. and Kenneth R. French, 1997. Industry costs of equity. Journal of Financial Economics, Vol. 43, pp. 153-193. Fama, Eugene F. and Kenneth R. French, 2008. Dissecting anomalies. The Journal of Finance, Vol. 63, No. 4, pp. 1653-1678. Faulkender, Michael and Rong Wang, 2006. Corporate financial policy and the value of cash. The Journal of Finance, Vol. 61, No. 4, pp. 1957-1990. Fazzari, Steven M., R. Glenn Hubbard, and Bruce C. Petersen, 1988. Financing constraints and corporate investment. Brookings Papers on Economic Activity, Vol. 1988, No. 1, pp. 141-206. Fazzari, Steven M., R. Glenn Hubbard, and Bruce C. Petersen, 2000. Investment-cash flow sensitivities are useful: A comment on Kaplan and Zingales. The Quarterly Journal of Economics, Vol. 115, No. 2, pp. 695-705. Fazzari, Steven M. and Bruce C. Petersen, 1993. Working capital and fixed invest- ment: New evidence on financing constraints. The RAND Journal of Economics, Vol. 24, No. 3, pp. 328-342. Fisher, Lawrence and Daniel Weaver, 1992. Dealing with short-term anomalies in the relative prices of securities, I: Constructing an unbiased equally weighted investment performance index and estimating the standard deviation of relative “errors”. Working paper, Rutgers University.

146 Fisher, Lawrence, Daniel Weaver, and Gwen Webb, 2010. Removing biases in com- puted returns. Review of Quantitative Finance and Accounting, Vol. 35, No. 2, pp. 137-161.

Fresard, Laurent, 2010. Financial strength and product market behavior: The real effects of corporate cash holdings. The Journal of Finance, Vol. 65, No. 3, pp. 1097-1122.

Gamba, Andrea and Alexander Triantis, 2008. The value of financial flexibility. The Journal of Finance, Vol. 63, No. 5, pp. 2263-2296.

Gantchev, Nickolay, 2013. The costs of shareholder activism: Evidence from a se- quential decision model. Journal of Financial Economics, Vol. 107, pp. 610-631.

Gantchev, Nickolay, Oleg Gredil, and Chotibhak Jotikasthira, 2015. Governance un- der the gun: Spillover effects of hedge fund activism. Working Paper.

Gaspar, Jos´e-Miguel and Massimo Massa, 2011. The role of commonality between CEO and divisional managers in internal capital markets. Journal of Financial and Quantitative Analysis, Vol. 46, No. 3, pp. 841-869.

Gertner, Robert H., Eric Powers, and David S. Scharfstein, 2002. Learning about internal capital markets from corporate spin-offs. The Journal of Finance, Vol. 57, No. 6, pp. 2479-2506.

Gillan, Stuart L. and Laura T. Starks, 2000. Corporate governance proposals and shareholder activism: The role of institutional investors. Journal of Financial Economics, Vol. 57, pp. 275-305.

Glaser, Markus, Florencio Lopez-de-Silanes, and Zacharias Sautner, 2013. Opening the black box: Internal capital markets and managerial power. The Journal of Finance, Vol. 68, No. 4, pp. 1577-1631.

Gomes, Joao and Dmitry Livdan, 2004. Optimal diversification: Reconciling theory and evidence. The Journal of Finance, Vol. 59, No. 2, pp. 507-535.

Graham, John R., Michael L. Lemmon, and Jack G. Wolf, 2002. Does corporate diversification destroy value? The Journal of Finance, Vol. 57, No. 2, pp. 695- 720.

Griffin, John M., Patrick J. Kelly, and Federico Nardari, 2010. Do market efficiency measures yield correct inferences? A comparison of developed and emerging markets. Review of Financial Studies, Vol. 23, No. 8, pp. 3225-3277.

Grinblatt, Mark and Tobias J. Moskowitz, 2004. Predicting stock price movements from past returns: The role of consistency and tax-loss selling. Journal of Finan- cial Economics, Vol. 71, pp. 541-579.

147 Hadlock, Charles J. and Joshua R. Pierce, 2010. New evidence on measuring financial constraints: Moving beyond the KZ index. Review of Financial Studies, Vol. 23, No. 5, pp. 1909-1940. Hartzell, Jay C. and Laura T. Starks, 2003. Institutional investors and executive compensation. The Journal of Finance, Vol. 58, No. 6, pp. 2351-2374. Hasbrouck, Joel, 2009. Trading costs and returns for U.S. equities: Estimating effec- tive costs from daily data. The Journal of Finance, Vol. 64, No. 3, pp. 1445-1477. Haugen, Robert A. and Nardin L. Baker, 1996. Commonality in the determinants of expected stock returns. Journal of Financial Economics, Vol. 41, pp. 401-439. Haushalter, David, Sandy Klasa, and William F. Maxwell, 2007. The influence of product market dynamics on a firm’s cash holdings and hedging behavior. Journal of Financial Economics, Vol. 84, pp. 797-825. He, Hua and David M. Modest, 1995. Market frictions and consumption-based asset pricing. Journal of Political Economy, Vol. 103, No. 1, pp. 94-117. Heaton, John and Deborah J. Lucas, 1996. Evaluating the effects of incomplete mar- kets on risk sharing and asset pricing. Journal of Political Economy, Vol. 104, No. 3, pp. 443-487. Hoberg, Gerard, Gordon Phillips, and Nagpurnanand Prabhala, 2014. Product mar- ket threats, payouts, and financial flexibility. The Journal of Finance, Vol. 69, No. 1, pp. 293-324. Hou, Kewei, G. Andrew Karolyi, and Bong-Chan Kho, 2011. What factors drive global stock returns? Review of Financial Studies, Vol. 24, No. 8, pp. 2527-2574. Hou, Kewei and Tobias J. Moskowitz, 2005. Market frictions, price delay, and the cross-section of expected returns. Review of Financial Studies, Vol. 18, No. 3, pp. 981-1020. Hou, Kewei, Chen Xue, and Lu Zhang, 2015. Digesting anomalies: An investment approach. Review of Financial Studies, Vol. 28, No. 3, pp. 650-705. Hou, Kewei, Chen Xue, and Lu Zhang, 2017. A comparison of new factor models. Working Paper, The Ohio State University. Hribar, Paul and Daniel W. Collins, 2002. Errors in estimating accruals: Implications for empirical research. Journal of Accounting Research, Vol. 40, No. 1, pp. 105- 134. Hund, John, Donald Monk, and Sheri Tice, 2010. Uncertainty about average prof- itability and the diversification discount. Journal of Financial Economics, Vol. 96, pp. 463-484.

148 Imbens, Guido and Karthik Kalyanaraman, 2012. Optimal bandwidth choice for the regression discontinuity estimator. Review of Economic Studies, Vol. 79, No. 3, pp. 933-959. Ince, Ozgur S. and R. Burt Porter, 2006. Individual equity return data from Thomson Datastream: Handle with care! Journal of Financial Research, Vol. 29, No. 4, pp. 463-479. Jegadeesh, Narasimhan and Sheridan Titman, 1993. Returns to buying winners and selling losers: Implications for stock market efficiency. The Journal of Finance, Vol. 48, No. 1, pp. 65-91. Jegadeesh, Narasimhan and Sheridan Titman, 2001. Profitability of momentum strategies: An evaluation of alternative explanations. The Journal of Finance, Vol. 56, No. 2, pp. 699-720. Jones, Charles M. and Owen A. Lamont, 2002. Short-sale constraints and stock re- turns. Journal of Financial Economics, Vol. 66, pp. 207-239. Kaplan, Steven N. and Luigi Zingales, 1997. Do investment-cash flow sensitivities provide useful measures of financing constraints? The Quarterly Journal of Eco- nomics, Vol. 112, No. 1, pp. 169-215. Karpoff, Jonathan M., Paul H. Malatesta, and Ralph A. Walkling, 1996. Corporate governance and shareholder initiatives: Empirical evidence. Journal of Financial Economics, Vol. 42, pp. 365-395. Keim, Donald B., 1983. Size-related anomalies and stock return seasonality: Further empirical evidence. Journal of Financial Economics, Vol. 12, pp. 13-32. Keynes, John Maynard, 1936. The general theory of employment, interest and money. Harcourt Brace, London. Klein, April and Emanuel Zur, 2009. Entrepreneurial shareholder activism: Hedge funds and other private investors. The Journal of Finance, Vol. 64, No. 1, pp. 187-229. Lakonishok, Josef and Seymour Smidt, 1984. Volume and turn-of-the-year behavior. Journal of Financial Economics, Vol. 13, pp. 435-455. Lamont, Owen, Christopher Polk, and Jes´usSa´a-Requejo,2001. Financial constraints and stock returns. Review of Financial Studies, Vol. 14, No. 2, pp. 529-554. Lang, Larry H.P. and Ren´eM. Stulz, 1994. Tobin’s q, corporate diversification, and firm performance. Journal of Political Economy, Vol. 102, No. 6, pp. 1248-1280. Lee, Kuan-Hui, 2011. The world price of liquidity risk. Journal of Financial Eco- nomics, Vol. 99, pp. 136-161.

149 Lesmond, David A., Joseph P. Ogden, and Charles A. Trzcinka, 1999. A new estimate of transaction costs. Review of Financial Studies, Vol. 12, No. 5, pp. 1113-1141.

Liu, Weimin, 2006. A liquidity-augmented capital asset pricing model. Journal of Financial Economics, Vol. 82, pp. 631-671.

Maksimovic, Vojislav, 1988. Capital structure in repeated oligopolies. The RAND Journal of Economics, Vol. 19, No. 3, pp. 389-407.

Maksimovic, Vojislav and Gordon Phillips, 2002. Do conglomerate firms allocate re- sources inefficiently across industries? Theory and evidence. The Journal of Finance, Vol. 57, No. 2, pp. 721-767.

Maksimovic, Vojislav and Gordon Phillips, 2013. Conglomerate firms, internal capital markets, and the theory of the firm. Annual Review of Financial Economics, Vol. 5, pp. 225-244.

Mansi, Sattar A. and David M. Reeb, 2002. Corporate diversification: What gets discounted? The Journal of Finance, Vol. 57, No. 5, pp. 2167-2183.

Modigliani, Franco and Merton H. Miller, 1958. The cost of capital, corporation finance, and the theory of investment. The American Economic Review, Vol. 48, No. 3, pp. 261-297.

Opler, Tim, Lee Pinkowitz, Ren´eM. Stulz, and Rohan Williamson, 1999. The de- terminants and implications of corporate cash holdings. Journal of Financial Economics, Vol. 52, No. 1, pp. 3-46.

Ozbas, Oguzhan and David S. Scharfstein, 2010. Evidence on the dark side of internal capital markets. Review of Financial Studies, Vol. 23, No. 2, pp. 581-599.

Parise, Gianpaolo, 2017. Threat of entry and debt maturity: Evidence from airlines. Journal of Financial Economics, Forthcoming.

Pastor, Lubos and Robert F. Stambaugh, 2003. Liquidity risk and expected stock returns. Journal of Political Economy, Vol. 111, No. 3, pp. 642-685.

Phillips, Gordon M., 1995. Increased debt and industry product markets: An empir- ical analysis. Journal of Financial Economics, Vol. 37, pp. 189-238.

Rajan, Raghuram, Henri Servaes, and Luigi Zingales, 2000. The cost of diversity: The diversification discount and inefficient investment. The Journal of Finance, Vol. 55, No. 1, pp. 35-80.

Roll, Richard, 1983. Vas ist das? The turn-of-the-year effect and the return premia of small firms. The Journal of Portfolio Management, Vol. 9, No. 2, pp. 18-28.

150 Scharfstein, David S. and Jeremy C. Stein, 2000. The dark side of internal capi- tal markets: Divisional rent-seeking and inefficient investment. The Journal of Finance, Vol. 55, No. 6, pp. 2537-2564.

Seru, Amit, 2014. Firm boundaries matter: Evidence from conglomerates and R&D activity. Journal of Financial Economics, Vol. 111, pp. 381-405.

Shin, Hyun-Han and Ren´eM. Stulz, 1998. Are internal capital markets efficient? The Quarterly Journal of Economics, Vol. 113, No. 2, pp. 531-552.

Sloan, Richard G., 1996. Do stock prices fully reflect information in accruals and cash flows about future earnings? Accounting Review, Vol. 71, No. 3, pp. 289-315.

Smith, Michael P., 1996. Shareholder activism by institutional investors: Evidence from CalPERS. The Journal of Finance, Vol. 51, No. 1, pp. 227-252.

Snider, Connan and Jonathan W. Williams, 2015. Barriers to entry in the airline industry: A multidimensional regression-discontinuity analysis of AIR-21. Review of Economics and Statistics, Vol. 97, No. 5, pp. 1002-1022.

Telser, Lester G., 1966. Cutthroat competition and the long purse. Journal of Law and Economics, Vol. 9, pp. 259-277.

Villalonga, Bel´en, 2004. Diversification discount or premium? New evidence from the Business Information Tracking Series. The Journal of Finance, Vol. 59, No. 2, pp. 479-506.

Wahal, Sunil, 1996. Pension fund activism and firm performance. Journal of Financial and Quantitative Analysis, Vol. 31, No. 1, pp. 1-23.

Weiss, Lawrence A. and Karen H. Wruck, 1998, Information problems, conflicts of interest, and asset stripping: Chapter 11’s failure in the case of Eastern Airlines. Journal of Financial Economics, Vol. 48, pp. 55-97.

Whited, Toni M., 2001. Is it inefficient investment that causes the diversification discount? The Journal of Finance, Vol. 56, No. 5, pp. 1667-1691.

Whited, Toni M. and Guojun Wu, 2006, Financial constraints risk. Review of Finan- cial Studies, Vol. 19, No. 2, pp. 531-559

151 Appendix A: Additional Figures and Tables to Chapter 1

This section is apportioned to the presentation of additional figures and tables that serve as addenda to the main analysis of the paper. They mostly pertain to side discussions (e.g. footnotes) within the text.

152 LCC Firm Passenger Passenger Number of Market Cash ST Debt LT Debt MC AIR21 Nonstop Share Share Rivals Size Firm Size Cash 1.00 0.08 -0.05 -0.09 -0.02 0.02 0.02 -0.05 0.00 -0.02 -0.45 ST Debt 0.08 1.00 0.24 -0.05 0.00 0.00 -0.04 -0.09 -0.01 0.02 -0.15 LT Debt -0.05 0.24 1.00 -0.05 0.00 -0.02 -0.09 -0.10 0.02 0.06 0.01 MC -0.09 -0.05 -0.05 1.00 0.06 -0.10 -0.02 0.02 -0.07 -0.19 0.59 AIR21 -0.02 0.00 0.00 0.06 1.00 0.04 -0.05 0.11 -0.20 -0.04 0.02 Nonstop 0.02 0.00 -0.02 -0.10 0.04 1.00 0.00 0.12 -0.07 0.09 -0.07 LCC Passenger Share 0.02 -0.04 -0.09 -0.02 -0.05 0.00 1.00 0.04 -0.05 -0.09 -0.11 Firm Passenger Share -0.05 -0.09 -0.10 0.02 0.11 0.12 0.04 1.00 -0.44 -0.22 0.11 Number of Rivals 0.00 -0.01 0.02 -0.07 -0.20 -0.07 -0.05 -0.44 1.00 0.45 -0.08 Market Size -0.02 0.02 0.06 -0.19 -0.04 0.09 -0.09 -0.22 0.45 1.00 -0.12 Firm Size -0.45 -0.15 0.01 0.59 0.02 -0.07 -0.11 0.11 -0.08 -0.12 1.00

153

Figure A.1: Correlations Between Explanatory and Control Variables

This figure tabulates the correlations between various explanatory and control variables that are included in the triple-difference regressions. Cash, ST Debt, and LT Debt are excess financial positions (scaled by total assets) of a firm relative to the average of all rival firms in the market as described in Section 1.3, MC is multimarket contact which captures the average number of markets jointly contested by the firm and its rivals in the market as described in Section 1.3, AIR21 equals 1 if the top-two airline passenger share at the origin airport of the market 2 years prior to the beginning of the present year exceeds 50% and equals 0 otherwise, Nonstop is an indicator for whether the service provided by the firm is a direct flight, LCC Passenger Share is the passenger share of low cost carriers (LCCs) in the market, Firm Passenger Share is the firm’s passenger share in the market, Number of Rivals is the number of rival firms excluding the firm itself in the market, Market Size is the log of total dollar revenues generated by all firms in the market, and Firm Size is the log of the firm’s total assets. Except for AIR21, all variables are lagged by 1 year. Table A.1: Passenger-Weighted Relative-to-Rival Cash and Multimarket Contact

This table presents results from triple-difference regressions of 12-quarter price growth differentials (∆∆36Log(Market Fares)) on an AIR-21 coverage indicator (AIR21), ex-ante passenger-weighted relative-to-rival cash holdings (Cash), ex-ante passenger-weighted multimarket contact (MC), and their interactions. Relative-to-rival cash and multimarket contact are constructed as described in Section 1.3, except that rivals are weighted by their passenger boardings in the market. The results are shown in five columns where the triple difference regression is run using progressively narrower sample windows surrounding the 50% AIR-21 treatment cutoff: (1) full sample, (2) 15% above and below the 50% cutoff, (3) 10% above and below the 50% cutoff, (4) 5% above and below the 50% cutoff, and (5) 2.5% above and below the 50% cutoff. In the next four columns, two sets of subsample analysis are conducted on a sample window 10% above and below the 50% AIR-21 treatment cutoff: (1) low cost carriers (LCCs) vs. legacy airlines and (2) ex-ante market winners vs. laggards. Firm, market, and time fixed effects as well as control variables are included in all specifications as in previous analysis. Standard errors adjusted for clustering at firm and market levels are reported in parentheses (*** p<0.01, ** p<0.05, * p<0.1).

Variables Dependent Variable: ∆∆36Log(Market Fares) Type (10%) Market Share (10%) Full Sample 15% 10% 5% 2.50% LCC Legacy Winners Laggards

154 AIR21 × Cash -0.547 -1.494*** -1.708** -2.719*** -3.996** 0.050 -2.012** -2.683** -0.608 (0.530) (0.514) (0.660) (0.753) (1.676) (0.720) (0.759) (1.134) (1.187)

AIR21 × MC × Cash 0.105 0.285** 0.309** 0.550*** 0.802* -0.141 0.344** 0.454* 0.078 (0.109) (0.109) (0.133) (0.163) (0.389) (0.125) (0.144) (0.226) (0.235)

MC -0.043*** -0.042*** -0.042** -0.047 -0.040 -0.122** -0.042 -0.036 -0.080*** (0.012) (0.014) (0.018) (0.030) (0.032) (0.041) (0.025) (0.039) (0.026)

Cash 1.220*** 1.377*** 1.250** 2.029*** 3.089** 0.843 1.348* 1.533* -0.092 (0.321) (0.363) (0.456) (0.681) (1.249) (0.724) (0.722) (0.792) (0.750)

MC × Cash -0.273*** -0.313*** -0.302*** -0.463*** -0.695** -0.109 -0.330** -0.330** -0.043 (0.067) (0.078) (0.089) (0.143) (0.312) (0.177) (0.129) (0.149) (0.153)

AIR21 0.031 0.065 0.126 -0.268* -0.344* -0.394** 0.096 0.117 0.021 (0.074) (0.091) (0.106) (0.138) (0.169) (0.122) (0.139) (0.217) (0.191)

AIR21 × MC -0.016 -0.023 -0.031 0.046 0.044 0.080** -0.026 -0.041 -0.003 (0.014) (0.016) (0.020) (0.028) (0.036) (0.025) (0.028) (0.038) (0.039)

Observations 15,958 9,097 6,047 2,828 1,701 1,773 4,075 2,887 2,858 Firms 19 19 19 19 18 7 12 16 18 Markets 1,941 1,404 1,036 561 394 470 818 795 603 Firm / Market / Time FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Controls Yes Yes Yes Yes Yes Yes Yes Yes Yes R-squared 0.227 0.272 0.303 0.351 0.401 0.412 0.329 0.382 0.356 Table A.2: Key Variable Statistics for Select Subsamples This table presents summary statistics of airline price growth differentials, relative-to-rival cash, and multimarket contact over the sample period 2001Q1 to 2014Q4 for a number of select subsamples. Variables are computed as detailed in Section 1.3. Statistics for each of these variables are shown across four subsamples: (1) sample window 15% above and below the 50% AIR-21 treatment cutoff, (2) sample window 5% above and below the 50% AIR-21 treatment cutoff (for which I report separately for default 1-year lagged, 3-year lagged, and 4-year lagged relative-to-rival cash), (3) legacy airlines in the sample window 2.5% above and below the 50% AIR-21 treatment cutoff, (4) and ex-ante market dominators (with ex-ante market share equal to or larger than 45%) in the sample window 15% above and below the 50% AIR-21 treatment cutoff.

Panel A. Price Growth Differentials Subsample Mean Std. Dev Min Med Max 15% above and below treatment cutoff 4.82 32.26 -197.99 4.47 184.24 5% above and below treatment cutoff 4.93 30.48 -133.37 3.26 154.45 Legacy & 2.5% above and below treatment cutoff 7.10 36.17 -120.39 5.03 145.47 Dominator & 15% above and below treatment cutoff 6.42 29.55 -107.74 5.37 177.52

Panel B. Relative-to-Rival Cash Subsample Mean Std. Dev Min Med Max 15% above and below treatment cutoff 0.30 8.55 -24.83 0.18 27.03 5% above and below treatment cutoff (1-year Lagged) 0.26 8.93 -24.74 -0.08 27.38 5% above and below treatment cutoff (3-year Lagged) 0.09 7.69 -21.04 0.17 21.97 5% above and below treatment cutoff (4-year Lagged) -0.42 7.62 -21.11 -0.58 21.31 Legacy & 2.5% above and below treatment cutoff -1.35 8.42 -21.44 -1.12 19.10 Dominator & 15% above and below treatment cutoff 0.23 8.49 -24.25 0.26 20.56

Panel C. Multimarket Contact Subsample Mean Std. Dev Min Med Max 15% above and below treatment cutoff 200.65 101.89 11.88 202.68 414.43 5% above and below treatment cutoff 199.16 99.99 12.10 204.00 422.69 Legacy & 2.5% above and below treatment cutoff 199.50 95.31 24.95 199.99 422.69 Dominator & 15% above and below treatment cutoff 210.38 108.81 12.88 219.45 411.04

155 Table A.3: Alternative Fixed Effects and Standard Error Clustering This table presents results from triple-difference regressions of 12-quarter price growth differentials (∆∆36Log(Market Fares)) on an AIR-21 coverage indicator (AIR21), ex-ante multimarket contact (MC), ex-ante relative-to-rival cash holdings (Cash), and their interactions. Panel A reports results controlling for firm-by-market fixed effects with standard errors adjusted for clustering at firm and market levels. Panel B reports results controlling for firm and market fixed effects separately, with standard errors adjusted for a variety of clustering schemes: (1) Firm and market (baseline), (2) firm-by-time and market, (3) firm-by-time, (4) firm and time, (5) market and time levels. Control variables are included as in previous analysis. Results are shown in five columns where triple difference regressions are run using progressively narrower sample windows surrounding the 50% AIR-21 treatment cutoff: (1) full sample, (2) 15% above and below the 50% cutoff, (3) 10% above and below the 50% cutoff, (4) 5% above and below the 50% cutoff, and (5) 2.5% above and below the 50% cutoff. Standard errors are reported in parentheses (*** p<0.01, ** p<0.05, * p<0.1).

Dependent Variable: ∆∆36Log(Market Fares) Panel A. Within Firm-Market Effects Full Sample 15% 10% 5% 2.50%

AIR21 × Cash -0.717 -2.429*** -2.288** -3.692*** -4.014 (0.701) (0.570) (0.843) (0.855) (4.014)

AIR21 × MC × Cash 0.116 0.424*** 0.357** 0.743*** 0.814 (0.132) (0.107) (0.164) (0.191) (0.850)

Observations 15,018 8,323 5,412 2,492 1,435 Firm-by-Market / Time FE Yes Yes Yes Yes Yes R-squared 0.306 0.362 0.379 0.419 0.487

Panel B. Alternative Clustering Schemes Full Sample 15% 10% 5% 2.50%

AIR21 × Cash -0.518 -1.776 -1.942 -2.967 -3.628

Firm and Market (Baseline) (0.543) (0.488)*** (0.665)*** (0.770)*** (1.706)** Firm-by-Time and Market (0.528) (0.650)*** (0.790)** (0.891)*** (1.647)** Firm-by-Time (0.510) (0.608)*** (0.760)** (0.813)*** (1.541)** Firm and Time (0.612) (0.690)** (0.711)** (0.844)*** (1.353)** Market and Time (0.605) (0.829)* (0.849)** (0.932)*** (1.172)**

AIR21 × MC × Cash 0.092 0.333 0.348 0.590 0.713

Firm and Market (Baseline) (0.110) (0.103)*** (0.132)** (0.159)*** (0.382)* Firm-by-Time and Market (0.111) (0.138)** (0.166)** (0.185)*** (0.364)* Firm-by-Time (0.109) (0.132)** (0.162)** (0.171)*** (0.347)** Firm and Time (0.131) (0.148)** (0.146)** (0.172)*** (0.314)** Market and Time (0.131) (0.172)* (0.176)* (0.189)** (0.264)**

Observations 15,958 9,097 6,047 2,828 1,701 Firm / Market / Time FE Yes Yes Yes Yes Yes R-squared 0.226 0.271 0.301 0.349 0.401

156 Table A.4: Difference of Means: Control vs. Treated Groups This table reports variable means of control (AIR21=0) and treatment (AIR21=1) observations, as well as p-values from t-tests of whether their means are equal (*** p<0.01, ** p<0.05, * p<0.1). This is done for the full sample as well as for a subsample where the top-two airline passenger share is within 15% of the 50% AIR-21 treatment cutoff (35% to 65%). Firm level variables include Cash, Short-Term Debt, Long-Term Debt, Total Debt, all scaled by total assets, and firm Size (log $ millions). Airport level variables include Market Size (log of $ revenues earned by all airlines serving markets originating from airport), average Distance of routes originating from airport (log miles), average Price Level of markets originating from airport (log $), Price Level charged by LCCs (log $), Price Level charged by Legacy airlines (log $), LCC Passenger Share, number of Passengers (log), and Number of Routes (log). Market level variables include Market Size, Distance, Price Level, Price Level LCC, Price Level Legacy, LCC Passenger Share, and number of Passengers, all measured analogously to airport level variables but at the market level. Firm-market level variables include Relative-to-Rival Cash/Assets, Short-Term Debt/Assets, Long-Term Debt/Assets, Multimarket Contact (log), all constructed as described in Section 1.3, Price Level charged by airline (log $), Price Growth during the previous 12 and 6 quarters (%), number of Passengers boarded by airline (log), airline’s Passenger Share in market, and Number of Rivals. All variables are lagged by 1 year.

Full 15% Variable Control Treated p-val Control Treated p-val Firm Level Cash / Asset 0.18 0.17 (0.44) 0.18 0.17 (0.39) ST Debt / Asset 0.04 0.04 (0.52) 0.04 0.04 (0.89) LT Debt / Asset 0.29 0.29 (0.69) 0.30 0.29 (0.57) Tot Debt / Asset 0.33 0.33 (0.72) 0.34 0.33 (0.48) Size (log $ millions) 15.00 14.88 (0.54) 14.99 14.94 (0.81) Airport Level Market Size (log $) 16.35 16.33 (0.77) 16.28 16.20 (0.48) Distance (log miles) 7.21 7.15 (0.06) * 7.21 7.21 (0.95) Price Level (log $) 5.27 5.31 (0.05) * 5.28 5.30 (0.31) Price Level LCC (log $) 5.06 5.06 (0.61) 5.06 5.07 (0.58) Price Level Legacy (log $) 5.39 5.42 (0.08) * 5.39 5.43 (0.05) * LCC Passenger Share 0.39 0.40 (0.80) 0.39 0.41 (0.44) Passengers (log) 11.08 11.01 (0.47) 11.00 10.89 (0.34) Num of Routes (log) 3.67 3.67 (0.93) 3.65 3.62 (0.62) Market Level Market Size (log $) 12.77 12.77 (0.79) 12.72 12.64 (0.00) *** Distance (log miles) 7.09 6.99 (0.00) *** 7.09 7.03 (0.00) *** Price Level (log $) 5.41 5.46 (0.00) *** 5.42 5.43 (0.33) Price Level LCC (log $) 5.20 5.20 (0.81) 5.21 5.22 (0.28) Price Level Legacy (log $) 5.38 5.42 (0.00) *** 5.39 5.40 (0.20) LCC Passenger Share 0.31 0.28 (0.00) *** 0.29 0.32 (0.00) *** Passengers (log) 7.42 7.36 (0.01) ** 7.36 7.26 (0.00) *** Firm-Market Level Cash / Asset (Relative-to-Rival) 0.00 0.00 (0.21) 0.00 0.00 (0.12) ST Debt / Asset (Relative-to-Rival) 0.00 0.00 (0.50) 0.00 0.00 (0.02) ** LT Debt / Asset (Relative-to-Rival) -0.01 -0.01 (0.39) -0.01 -0.01 (0.27) Multimarket Contact (log) 4.98 5.05 (0.00) *** 5.00 5.12 (0.00) *** Price Level (log $) 5.40 5.45 (0.00) *** 5.41 5.42 (0.16) Price Growth - Prev. 12Q (%) 1.72 -0.04 (0.00) *** 1.72 1.11 (0.15) Price Growth - Prev. 6Q (%) 3.46 2.25 (0.00) *** 3.25 2.41 (0.01) ** Passengers (log) 5.18 5.20 (0.32) 5.14 5.18 (0.12) Passenger Share 0.36 0.41 (0.00) *** 0.36 0.40 (0.00) *** Number of Rivals 2.62 2.05 (0.00) *** 2.52 2.14 (0.00) ***

157 Appendix B: Data and Variable Descriptions to Chapter 2

This section describes the construction of variables used in this study. To miti- gate the influence of outliers, all continuous variables are winsorized at the 1st and

99th percentiles, except for segment Tobin’s Q which is used to construct categorical variables.

B.1 Variables

• HFA: Dummy variable that equals 1 if the date of the first Schedule 13D filing

with the firm as the investment target was at least 1 year ago, and 0 otherwise.

• INVS: Segment level gross capital expenditure (item CAPXS) divided by

lagged firm total assets (item AT).

• CF: Segment level cash flow, computed as operating income before depreciation

(item OIBDPS) or operating profit (item OPS) plus depreciation (item DPS),

as available, divided by lagged firm total assets (item AT).

• Other CF: Sum of cash flows (item OIBDPS or items OPS plus DPS as

available) of all other segments in same firm, divided by lagged firm total assets

(item AT).

158 • Segment Q: Segment level Tobin’s Q is the median Tobin’s Q of single-segment

firms in the same two-digit SIC code industry as the segment. Tobin’s Q of a

single-segment firm is computed as the ratio of firm value (defined as the market

value of equity (item CSHO multiplied by item PRCC F) plus the book value

of total assets (item AT) minus the book value of equity (item CEQ plus item

TXDB)) to the book value of total assets (item AT).

• High Q: Dummy variable that equals 1 if the segment’s Tobin’s Q is

– highest among all segments in the same firm.

– above the median Tobin’s Q of segments in the same firm.

– above the average Tobin’s Q of segments in the same firm.

– lowest among all segments in the same firm.

• Control Variables: Segment level Tobin’s Q, segment sales (item SALES)

growth rate, firm level asset size (log of item AT), firm level profitabiltiy (ROA,

defined as income before extraordinary items (item IB) divided by lagged total

assets (item AT)), firm level cash holdings (item CHE divided by item AT),

firm level long-term leverage (item DLTT divded by item AT), all lagged by 1

year.

B.2 Data Screening

FASB No. 14 and SEC Regulation S-K require firms to report audited footnote information for business segments whose sales, assets, or profits comprise more than

10% of the firm’s consolidated totals. In June of 1997, FASB No. 14 was superseded

159 by FASB No. 131, under which firms are required to report such segment data inso-

far as “it is used internally for evaluating segment performance and deciding how to

allocate resources to segments”. The Compustat segment database reports segment

information based on this requirement. To ensure that the reporting requirement

change does not affect the results of the paper, I redo the analysis using the sample

period beginning with the fiscal year 1998 so that all variables, including the lagged

ones, use data strictly after the change occurred. The results are virtually unchanged.

To construct the variables described above, I begin by following Shin and Stulz

(1998) and require segments to contain complete information on net sales (item

SALES), identifiable total assets (item IAS), capital expenditures (item CAPXS),

operating profit (loss) (item OPS), depreciation (item DPS), and SIC code. I exclude

firms with financial segments (SIC codes between 6000 and 6999), since applying

Tobin’s Q as a measure of investment opportunities may be problematic in these industries.

There are a number of widely recognized issues with the Compustat segment database. For example, Compustat reports only up to ten segments, meaning smaller segments may be neglected. Moreover, firms may choose to allocate their financial reporting across segments with some discretion. As a result, firms may not fully allocate accounting items across the reported segments. To address this problem, I follow Berger and Ofek(1995), Billett and Mauer(2003), and Seru(2014) and require the sum of segment sales (assets) to be within 1% (25%) of firm totals, after which

I apply a multiple to explicitly allocate unallocated sales, assets, capital expendi- ture, and cash flow. Another important problem is that firms may reorganize their segments over time and this may distort the identification of particular segments.

160 To minimize any bias arising from this issue, I take cue from Shin and Stulz(1998) and require the following ratios to be less than one: segment capital expenditure to segment assets, other segment capital expenditure to other segment assts, segment capital expenditure to firm total assets, other segment capital expenditure to firm total assets, segment cash flow to firm total assets, and other segment cash flow to

firm total assets.

To ensure that the sample of firms are truly diversified, I follow Shin and Stulz

(1998) and Billett and Mauer(2003) and require firms to have at least two seg- ments serving different two-digit SIC industries and further exclude firms in which the smallest and largest segments are in the same industry.

161 Appendix C: Addendum to Chapter 3

C.1 Adjusting FRIC for Serial Correlation in True Returns

Our empirical estimation produces negative FRIC values for some firm-months.

These negative estimates can be the result of sampling errors, or very small mi- crostructure frictions combined with positively autocorrelated true returns (as shown by equation (3.15) in Section 3.2.2). To evaluate whether true return autocorrela- tion, which is assumed to be zero in our baseline estimation, can impair our ability to cleanly identify the price impact of microstructure frictions, we conduct two addi- tional sets of analyses in this appendix.

We first remove firm-month observations where FRIC is negative, and repeat the portfolio sorts in Table 3.2 Panel A. To the extent that a negative FRIC value is caused by positively autocorrelated true returns, this ensures that the positive return autocorrelation does not drive our results. Panel A of Table C.1 show that the friction premium based on positive-only FRIC is significant both economically and statisti- cally, albeit at a slightly smaller magnitude compared with the unconstrained FRIC sorts.39 The average return spread between Deciles 10 and 1 is 1.05% (t = 3.61) per month for equal-weighted returns and 0.74% (t = 2.06) per month for value-weighted

39The smaller magnitude is not surprising as the positive-only filter eliminates a nontrivial number of firms with potentially small microstructure frictions.

162 returns. All of the factor model alphas, except the Fama and French(1993) three- factor alpha for value-weighted returns, are large and significant.

We also attempt to explicitly adjust FRIC for the effect of true return autocor- relation using equation (3.15). To do so, we need an estimate of the autocorrelation coefficient, ρi. We begin by comparing a three-period return relative to a two-period return, both ending at time t. Specifically, the expected three-period true return,

E [1 + ri,t−3,t], can be decomposed as follows.

E [1 + ri,t−3,t] = E{[1 + ri,t−3,t−1] [1 + ri,t−1,t]} (C.1) = Cov (ri,t−3,t−1, ri,t−1,t) + E [1 + ri,t−3,t−1] E [1 + ri,t−1,t]

Using the definition of the two-period return, expressed in terms of one-period returns, (1 + ri,t−3,t−1) = (1 + ri,t−3,t−2)(1 + ri,t−2,t−1), we can rewrite the covariance term in (C.1) as

Cov (ri,t−3,t−1, ri,t−1,t) =Cov (ri,t−3,t−2, ri,t−1,t) + Cov (ri,t−2,t−1, ri,t−1,t) (C.2) + Cov (ri,t−3,t−2 · ri,t−2,t−1, ri,t−1,t)

With the assumed autoregressive true return process, and the added assumption that Cov (ri,t−3,t−2 · ri,t−2,t−1, ri,t−1,t) = 0, this equation can be rewritten as

ρ2σ2 ρ σ2 i i i i Cov (ri,t−3,t−1, ri,t−1,t) = 2 + 2 (C.3) (1 − ρi ) (1 − ρi )

Analogous to the procedure for cancelling out the microstructure frictions using the two-period and the one-period observed returns as described in Section 3.2.2, we now have

h 2 i 1 + σ E [1 + ri,t−3,t] E [1 + rbi,t−3,t] ϕi,t−3 E [1 + ri,t−3,t] ≈ h i = E [1 + rbi,t−3,t−1] 1 + σ2 E [1 + r ] E [1 + ri,t−3,t−1] ϕi,t−3 i,t−3,t−1 (C.4) ρ2σ2 ρ σ2 i i i i = E [1 + ri,t−1,t] + 2 + 2 (1 − ρi ) (1 − ρi ) 163 and it then follows that

E [1 + r ] ρ2σ2 ρ σ2 E [1 + r ] − bi,t−3,t ≈ σ2 − i i − i i (C.5) bi,t−1,t ϕi,t−1 2 2 E [1 + rbi,t−3,t−1] (1 − ρi ) (1 − ρi )

ρ σ2 E[1+rbi,t−2,t] 2 i i Now recall from equation (3.15), E [1 + ri,t−1,t] − ≈ σ − . E[1+ri,t−2,t−1] ϕi,t−1 2 b b (1−ρi )

Taking the difference between equation (3.15), which we label ∆2,1, and equation

(C.5), which we label ∆3,2, we can obtain an estimate of the two-period autocovariance

of true returns. ρ2σ2 i i ∆2,1 − ∆3,2 ≈ 2 (C.6) (1 − ρi ) A similar derivation based on the four-period and the two-period return leads to

ρ2 (2 + ρ ) σ2 i i i ∆2,1 − ∆4,2 ≈ 2 (C.7) (1 − ρi )

The serial correlation coefficient, ρi, can then be obtained by dividing equation

(C.7) by equation (C.6).

∆2,1 − ∆4,2 ≈ 2 + ρi (C.8) ∆2,1 − ∆3,2 Using equation (C.8), we can in principle estimate the serial correlation coefficient

of true returns, ρi, by taking the ratio of the differences in FRIC estimated using alter-

native windows. The numerator is the baseline FRIC (estimated by dividing average

two-period observed return by average lagged one-period observed returns) minus the

FRIC based on average four-period observed return divided by average lagged two-

period observed returns, and the denominator is the baseline FRIC minus the FRIC

based on average three-period observed return divided by average lagged two-period

observed returns. Then, dividing equation (C.6) by the estimated serial correlation

coefficient ρi and adding that piece back to equation (3.15) effectively removes the

autocorrelation term, leaving the friction term σ2 only. ϕi,t−1

164 However, implementing this estimation at the firm level would introduce sampling

errors at multiple steps of the procedure. To mitigate this problem, we estimate equa-

tion (C.8) using FRIC-sorted portfolios instead. Specifically, each month we run a

firm-level pooled regression of ∆2,1 − ∆4,2 on ∆2,1 − ∆3,2 over the past year for each decile portfolio sorted by the baseline FRIC measure. This gives us one ρd estimate for each decile portfolio, d, which is then assigned to individual firms within that portfolio. We then adjust our firm-level FRIC measure for autocorrelation of true returns as described above using the portfolio-level ρd estimate.

Panel B of Table C.1 reports equal and value-weighted returns of decile portfo- lios sorted on ρd-adjusted FRIC. While the magnitudes are smaller than the baseline results, both the equal and value-weighted friction premiums are positive and statis- tically significant. The equal-weighted raw and factor-adjusted return spreads are all significant, reaching as high as 76 basis points per month. The value-weighted raw return spread, Carhart(1997) alpha, and q-factor alphas are all significant at the

10% level, and can be as large as 25 basis points per month.

Overall, these additional tests suggest that our main results are robust to con- trolling for the effect of true return autocorrelation on our FRIC measure. However, note that these alternative procedures are not without issues (truncating the sample, sampling errors). On balance, we believe that our baseline approach of assuming zero true return autocorrelation is the most efficient way of capturing the impact of microstructure frictions on returns.

165 Table C.1: Microstructure Frictions and the Cross-Section of Expected Stock Returns: Addressing Serial Correlation in True Returns

The equal- and value-weighted average monthly returns (%) of friction-sorted decile portfolios, their t-statistics (in parentheses), and the difference in returns between decile portfolios 10 (largest friction) and 1 (smallest friction) are reported over the period July, 1966, to June, 2013. Panel A reports the equal- and value-weighted results using only observations with positive FRIC values. Panel B reports the equal- and value-weighted results from sorting on ρd-adjusted FRIC as described in Appendix C.1. The last four columns in Panel A and Panel B report the intercepts, or alphas, from time-series regressions of the 10-1 average return spreads on the Fama and French(1993) three-factor model, Carhart(1997) four-factor model (which adds a momentum factor to Fama and French(1993) factors), a five-factor model that adds the Pastor and Stambaugh(2003) aggregate liquidity risk factor-mimicking portfolio to the Carhart(1997) four-factor model, and the Hou, Xue, and Zhang(2015), in short HXZ, q-factor model. Panel A: Returns on Positive-Only FRIC-sorted portfolios

166 (10-1) (10-1) (10-1) (10-1) Decile 1 2 3 4 5 6 7 8 9 10 (10-1) FF Carhart Pastor and HXZ portfolio 3-factor 4-factor Stambaugh q-factor 5-factor Equal-weighted 1.20 1.23 1.28 1.18 1.25 1.37 1.36 1.35 1.57 2.24 1.05 0.75 1.00 1.00 1.62 (5.95) (5.82) (5.87) (5.24) (5.23) (5.45) (5.04) (4.53) (4.78) (5.61) (3.61) (2.95) (3.93) (3.83) (6.25) Value-weighted 0.90 0.97 0.97 0.95 1.00 1.00 0.93 1.02 1.05 1.64 0.74 0.20 0.56 0.61 1.26 (4.80) (5.17) (4.91) (4.68) (4.38) (4.07) (3.44) (3.27) (3.08) (3.95) (2.06) (0.71) (1.96) (2.06) (4.27)

Panel B: Returns on ρd-Adjusted FRIC-sorted portfolios (10-1) (10-1) (10-1) (10-1) Decile 1 2 3 4 5 6 7 8 9 10 (10-1) FF Carhart Pastor and HXZ portfolio 3-factor 4-factor Stambaugh q-factor 5-factor Equal-weighted 1.01 1.13 1.18 1.25 1.25 1.24 1.24 1.30 1.43 1.69 0.68 0.61 0.65 0.64 0.76 (3.10) (3.91) (4.64) (5.66) (6.11) (6.06) (5.61) (5.22) (4.97) (4.90) (6.12) (5.65) (5.87) (5.63) (6.59) Value-weighted 0.67 0.94 0.97 0.97 1.02 0.94 0.82 0.85 0.73 0.91 0.23 0.16 0.24 0.22 0.25 (2.35) (3.82) (4.44) (4.90) (5.50) (5.08) (3.96) (3.69) (2.96) (3.10) (1.81) (1.19) (1.78) (1.61) (1.72) Table C.2: Fama-Macbeth Regressions (WLS) Robustness test of Table 3.3 via weighted least squares (WLS) instead of ordinary least squares (OLS), where each observed return is weighted by the lagged gross observed return on the same security, following Asparouhova, Bessembinder, and Kalcheva(2010, 2013). Panel A presents results from Fama and MacBeth(1973) monthly cross-sectional regressions of stock returns in excess of the one-month T-bill rate on our friction measure (FRIC), log of firm size (market capitalization), log of book-to-market equity (BE/ME), previous month’s return (month t − 1, Ret−1:−1 (%)), previous year’s return (from month t − 12 to t − 2, Ret−12:−2 (%)), previous three year’s return (from month t − 36 to t − 13, Ret−36:−13 (%)), log of previous month’s share price, profitability (equity income divided by total assets), asset growth (change in total assets divided by total assets), accrual (Prior to 1988, accruals are calculated using the balance sheet method as changes in non-cash current assets, 167 less changes in current liabilities excluding changes in short-term debt and changes in taxes payable, minus depreciation and amortization expenses. Starting in 1988, accruals are calculated using the cash flow statement method as the difference between earnings and cash flows from operations), idiosyncratic volatility (IVOL: standard deviation of residuals from Fama and French(1993) three-factor model using daily returns from previous month), and a host of alternative liquidity variables over the period July, 1966, to June, 2013. Liquidity variables include average daily turnover (average daily number of shares traded, divided by number of shares outstanding from prior year), volume (average daily dollar trading volume over the past year), Amihud(2002) illiquidity measure (daily absolute return divided by daily dollar trading volume, averaged over prior year), average daily effective bid-ask spread, Corwin and Schultz(2012) spread estimate (CS), zero return proportion measure (ZP: number of days with zero return divided by number of trading days from previous year), and Hasbrouck(2009) effective cost measure (C), each defined separately for NYSE/AMEX and Nasdaq traded firms. The time-series averages of the coefficient estimates and their associated time-series t-statistics (in parentheses) are reported in the style of Fama and MacBeth(1973). Panel B shows results from two kitchen sink regressions, where all control variables and liquidity measures are included in the regressions. The first kitchen sink specification excludes FRIC and the second specification includes all variables. Table C.2 - Continued. Panel A: Fama-MacBeth Regressions Panel B: Fama-MacBeth Regressions (Kitchen Sink) Liquidity Measures (1) (2) NYSE Nasdaq Turnover Volume Amihud BIDASK CS ZP C FRIC 4.01 /AMEX (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (4.59) log(Size) 0.00 -0.01 -0.05 0.09 -0.03 0.00 0.00 0.02 -0.02 -0.03 -0.01 Turnover 0.06 0.06 (-0.12) (-0.25) (-1.80) (1.36) (-0.93) (-0.04) (0.04) (0.49) (-0.75) (-0.88) (-0.37) (0.31) (0.29) log(BE/ME) 0.23 0.23 0.20 0.28 0.22 0.21 0.22 0.21 0.20 0.22 0.21 Volume -0.03 -0.02 (4.47) (4.39) (3.94) (4.67) (4.57) (4.57) (4.38) (4.34) (3.94) (4.88) (3.98) (-0.14) (-0.10) Ret−1:−1 -5.19 -5.12 -4.77 -3.97 -5.32 -5.35 -5.20 -5.26 -5.30 -5.20 -5.59 Amihud 8.22 7.83 (-16.26) (-16.07) (-12.32) (-11.60) (-17.01) (-17.18) (-16.13) (-16.89) (-16.16) (-17.36) (-16.76) (0.90) (0.85) Ret−12:−2 0.73 0.73 0.70 0.61 0.76 0.76 0.74 0.75 0.79 0.74 0.82 BIDASK -15.98 -16.22 (6.03) (6.09) (4.51) (5.61) (6.42) (6.47) (6.10) (6.43) (6.39) (6.58) (6.39) (-2.82) (-2.87) Ret -0.09 -0.09 -0.08 -0.05 -0.09 -0.09 -0.09 -0.10 -0.09 -0.10 -0.09 CS -5.73 -7.83

−36:−13 NYSE/AMEX (-2.03) (-2.04) (-1.43) (-1.38) (-2.16) (-2.17) (-2.03) (-2.46) (-2.16) (-2.27) (-1.91) (-1.15) (-1.58)

168 Log(Price) -0.32 -0.29 -0.14 -0.41 -0.26 -0.25 -0.25 -0.31 -0.24 -0.28 -0.18 ZP -2.88 -2.80 (-3.82) (-3.50) (-1.99) (-3.17) (-3.31) (-3.13) (-3.05) (-3.51) (-2.91) (-3.06) (-1.92) (-3.67) (-3.57) Profitability 1.23 1.24 1.62 0.63 1.03 1.04 1.08 1.15 1.12 1.24 1.31 C 33.99 25.04 (4.84) (4.85) (5.30) (2.77) (4.00) (4.03) (4.03) (4.76) (4.14) (4.90) (4.46) (3.52) (2.57) Asset growth -0.23 -0.22 -0.23 -0.25 -0.19 -0.19 -0.20 -0.20 -0.21 -0.22 -0.23 Turnover -0.16 -0.15 (-4.41) (-4.27) (-3.48) (-4.95) (-3.38) (-3.31) (-3.33) (-4.08) (-3.53) (-4.47) (-3.62) (-1.17) (-1.07) Accruals -0.96 -0.94 -1.17 -0.58 -0.98 -0.98 -0.98 -0.89 -1.03 -0.94 -1.14 Volume 0.15 0.13 (-6.69) (-6.57) (-5.50) (-2.82) (-6.56) (-6.56) (-6.49) (-6.31) (-6.66) (-6.65) (-7.17) (1.04) (0.95) IVOL -8.29 -9.93 -14.65 -7.88 -11.72 -11.54 -14.04 -6.59 -13.72 -9.43 -15.21 Amihud 145.05 147.57 (-3.27) (-3.82) (-4.60) (-2.78) (-5.48) (-5.47) (-5.36) (-2.94) (-5.25) (-4.24) (-5.62) (1.49) (1.50) FRIC 3.88 5.96 1.78 4.26 4.31 3.99 4.15 5.18 3.88 3.96 BIDASK -10.33 -12.04 (4.15) (4.70) (2.32) (4.83) (4.92) (4.50) (4.41) (5.86) (4.45) (4.06) (-1.64) (-1.87) Nasdaq -0.09 0.08 -0.81 -0.79 -0.76 -1.10 -0.90 NASDAQ CS -7.65 -7.88 (-0.18) (0.03) (-2.34) (-0.85) (-2.02) (-1.56) (-2.24) (-1.04) (-1.09) Liquidity -0.03 -0.02 27.39 -9.04 -7.11 -1.31 13.78 ZP 3.91 4.12 AMEX/NYSE (-0.49) (-0.27) (3.20) (-2.14) (-1.65) (-1.52) (1.90) (0.90) (0.95) Liquidity -1.04 -0.04 45.74 12.91 5.09 2.06 38.35 C 69.09 61.95 Nasdaq (-1.51) (-0.16) (0.35) (0.61) (0.73) (1.07) (1.90) (1.14) (1.02) Control Variables Yes Yes