Open Collections

UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

Essays in real estate finance : mortgage contract terms, pricing and behaviour Wetzel, Jacob Anders 2017

Your browser doesn't seem to have a PDF viewer, please download the PDF to view this item.

Item Metadata


24-ubc_2017_september_wetzel_jacob.pdf [ 2.49MB ]
JSON: 24-1.0354552.json
JSON-LD: 24-1.0354552-ld.json
RDF/XML (Pretty): 24-1.0354552-rdf.xml
RDF/JSON: 24-1.0354552-rdf.json
Turtle: 24-1.0354552-turtle.txt
N-Triples: 24-1.0354552-rdf-ntriples.txt
Original Record: 24-1.0354552-source.json
Full Text

Full Text

Essays in Real Estate Finance: Mortgage Contract Terms, Pricingand BehaviourbyJacob Anders WetzelB.A., University of California, Berkeley, 2002MSc, Oxford University, 2006MSc, The University of British Columbia, 2009A THESIS SUBMITTED IN PARTIAL FULFILLMENTOF THE REQUIREMENTS FOR THE DEGREE OFDoctor of PhilosophyinTHE FACULTY OF GRADUATE AND POSTDOCTORAL STUDIES(Business Administration)The University of British Columbia(Vancouver)September 2017© Jacob Anders Wetzel, 2017AbstractThis thesis is a collection of three essays in Real Estate Finance. The first essay examines thedeterminants of commercial mortgage contract terms. A cornerstone of finance theory is that riskand return should be positively related. However, existing empirical studies often find a negativerelationship between interest rates and risk terms in mortgage contracts. Previous studies havefound such results puzzling, and have surmised that they may arise because traditional modelsdo not explicitly account for the simultaneous determination of interest rates and loan terms. Wetherefore specify separate supply and demand equations for loanable funds, and then examinemortgage contracts as equilibrium outcomes of a multidimensional negotiation between borrowerand lender. Our empirical results reveal that borrowers and lenders individually require higherreturns for greater risk, as theory requires, but that this can produce apparently negative risk/returncorrelations in contract outcomes, as observed in the data. We demonstrate how various risk factorsimpact the simultaneous determination of equilibrium interest rates and loan terms.The second essay investigates adverse selection in the Home Equity Conversion Mortgages(“HECMs”). The pricing structure used by the Federal Housing Administration (“FHA”) does notreflect geographic or cyclical risk. Since HECM’s were disproportionately originated in the sandstates in the lead up to the 2008 financial crises, the majority of the loans originated between 2005-2007 were underwater. We ask whether borrowers adversely selected into HECM’s with the intentto exploit mispriced insurance? This appears unlikely: borrowers whose loans terminated withcredit limits greater than their homes are worth have been no likelier to exhaust credit than similarborrowers whose loans terminated with credit limits below collateral value.The third essay studies the effect on residential property prices arising from proximity to oilpipelines. The key contributions of the paper are to show that [1] the disamenity effects relatedto pipeline proximity are highly localized over very short distances [2] the magnitude of the effectsare sensitive to the land use of the pipeline easement. Our findings suggest a likely specificationbias in studies that use parametric measures of proximity to an environmental hazard.iiLay SummaryThis thesis is a collection of three essays in Real Estate Finance. Although the topics are diverse,they address fundamental questions related to mortgage contract terms, pricing and behaviour inresidential and commercial property markets. The first essay studies the determinants of contractterms in commercial mortgages. The second essay investigates strategic default by borrowers inHome Equity Conversion Mortgages. The third essay studies the effect on residential propertyprices arising from proximity to oil pipelines.iiiPrefaceComplying with UBC guidelines, I state here the role of each of the co-authors. The essay inchapter 2 is an unpublished paper that I co-authored with Professor Glen Donaldson (University ofBritish Columbia). The essay in chapter 3 is an unpublished paper that I co-authored with ProfessorTom Davidoff (University of British Columbia). The essay in chapter 4 is an unpublished paperthat I co-authored with Professor Tsur Somerville (University of British Columbia). In each ofthe co-authored projects, all authors worked together on all aspects of the paper. The developmentof each research project involved frequent meetings to discuss the identification of the researchquestion, the theoretical analysis, the data analysis, the empirical work, and the writing. On eachof these papers I was a full partner.ivTable of ContentsAbstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iiLay Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iiiPreface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ivTable of Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vList of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viiList of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viiiAcknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ix1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12 The Simultaneous Determination of Interest Rates and Loan Terms: Evidence fromthe Mortgage Market . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32.2 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92.3 The Risk-Return Puzzle and Endogeneity of Contract Terms . . . . . . . . . . . . 122.3.1 Spread Regressions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152.3.2 Multidimensional Tradeoffs Among Risk Variables . . . . . . . . . . . . . 172.4 A Simultaneous Equations Model of the Mortgage Market - Specification and Es-timation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192.4.1 Model Specification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 202.4.2 Empirical Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 252.5 Comparative Statics, Puzzle Resolutions, and the Simultaneous Determination ofLoan Terms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 272.5.1 Changing the Riskfree Rate . . . . . . . . . . . . . . . . . . . . . . . . . 292.5.2 Amortization Effects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 312.5.3 Solving the Loan-to-Value Puzzle . . . . . . . . . . . . . . . . . . . . . . 332.5.4 Solving the Maturity Puzzle . . . . . . . . . . . . . . . . . . . . . . . . . 372.5.5 Property Type . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 402.5.6 Originator Type . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 412.5.7 Simultaneous Shifts from Multiple Factors . . . . . . . . . . . . . . . . . 442.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46v3 Do Reverse Mortgage Borrowers Use Credit Ruthlessly? . . . . . . . . . . . . . . . 483.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 483.2 The Home Equity Conversion Mortgage (“HECM”) . . . . . . . . . . . . . . . . 553.2.1 HECM loan structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 553.2.2 The embedded put option . . . . . . . . . . . . . . . . . . . . . . . . . . . 573.3 Empirical Analysis of HECM Credit Lines . . . . . . . . . . . . . . . . . . . . . . 613.3.1 Adverse Selection into HECM on Put Option Value . . . . . . . . . . . . . 613.3.2 Estimation Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . 653.3.3 HECM Loan Microdata . . . . . . . . . . . . . . . . . . . . . . . . . . . 703.3.4 Samples of loans with and without in-the-money put options . . . . . . . . 803.3.5 Plausibility of Identifying Assumptions . . . . . . . . . . . . . . . . . . . 873.3.6 Credit Exhaustion Among Selected Borrowers . . . . . . . . . . . . . . . 943.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1074 Fear and Loathing of Oil Pipelines: Hunting for Disamenity Effects . . . . . . . . . 1134.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1134.2 Literature Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1164.3 Data and Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1234.3.1 Identification challenges and proposed tests . . . . . . . . . . . . . . . . . 1254.3.2 Description of sample and summary statistics . . . . . . . . . . . . . . . . 1294.4 Regression Specification and Results . . . . . . . . . . . . . . . . . . . . . . . . . 1334.5 Event Studies : Difference in Difference Regressions . . . . . . . . . . . . . . . . 1394.5.1 Effect of Spill . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1404.5.2 Effect of Expansion Announcement . . . . . . . . . . . . . . . . . . . . . 1424.6 Summary and Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1445 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147viList of TablesTable 2.1 Summary Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11Table 2.2 Information Regarding Specific Mortgages in Figure 2.1 . . . . . . . . . . . . . 13Table 2.3 Basic Regressions of Spread on Risk Indicators . . . . . . . . . . . . . . . . . 16Table 2.4 Estimated Supply and Demand Curves . . . . . . . . . . . . . . . . . . . . . . 26Table 2.5 Property Types by Originator . . . . . . . . . . . . . . . . . . . . . . . . . . . 42Table 3.1 Summary statistics of sample . . . . . . . . . . . . . . . . . . . . . . . . . . . 73Table 3.2 List of treated and comparison metropolitan areas . . . . . . . . . . . . . . . . 84Table 3.3 Characteristics of loans in treated and comparison metropolitan areas: . . . . . 85Table 3.4 OLS regressions of an indicator for loan termination . . . . . . . . . . . . . . 92Table 3.5 Panel logit regressions of credit “exhaustion” . . . . . . . . . . . . . . . . . . 94Table 3.6 OLS regressions of an indicator for near exhaustion of credit (95%+ use) . . . . 96Table 4.1 Descriptive statistics for all transactions . . . . . . . . . . . . . . . . . . . . . 130Table 4.2 Frequency counts on distance to pipeline easement and adjacency . . . . . . . . 132Table 4.3 Cross Tabulation of Adjacency and Distance Bands Measures . . . . . . . . . . 133Table 4.4 Baseline regression specifications with simple distance measures . . . . . . . . 134Table 4.5 Distance in Discrete Bands . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137Table 4.6 Distance in Adjacency Measures . . . . . . . . . . . . . . . . . . . . . . . . . 138Table 4.7 Difference in Difference Regressions: Effect of Oil Spill (250m Bands) . . . . . 141Table 4.8 Robustness Check : Westridge Oil Spill Falsifications Regressions . . . . . . . 142Table 4.9 Pipeline Expansion Announcement Event Study Regressions . . . . . . . . . . 143viiList of FiguresFigure 2.1 Loan Price vs Quantity from Mortgage Data . . . . . . . . . . . . . . . . . . . 13Figure 2.2 LTV , CapRate, Amortization and LTI . . . . . . . . . . . . . . . . . . . . . . 18Figure 2.3 Inverse Supply and Demand Curves - Base Case . . . . . . . . . . . . . . . . 28Figure 2.4 Comparative Static - Changing the Riskfree Rate . . . . . . . . . . . . . . . . 30Figure 2.5 Changing Amortization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32Figure 2.6 Changing Value-to-Income . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35Figure 2.7 Changing Maturity Term . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38Figure 2.8 Changing Property Type . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41Figure 3.1 HECM origination vs metropolitan area price change by origination quarter . . 62Figure 3.2 Average put “moneyness” by origination quarter . . . . . . . . . . . . . . . . 63Figure 3.3 Sand State origination share by origination quarter . . . . . . . . . . . . . . . 64Figure 3.4 Distribution of loans terminating “underwater” by quarter of origination . . . . 72Figure 3.5 Proportion of loans assigned to FHA by estimated Loan to Value . . . . . . . 76Figure 3.6 Comparison of proportion of loans assigned to FHA by originator . . . . . . . 77Figure 3.7 Comparison of proportion of loans assigned to FHA by metropolitan area . . . 79Figure 3.8 Distribution of HECM origination share and metro area price growth . . . . . . 82Figure 3.9 Distribution of credit use across Treated and Comparison metropolitan areas . . 101Figure 3.10 Distribution of credit use among borrowers in the “treated” metropolitan areas . 106Figure 4.1 Study area with Transactions . . . . . . . . . . . . . . . . . . . . . . . . . . . 125viiiAcknowledgmentsI wish to express my sincere gratitude to my thesis advisors and co-authors for their invaluableguidance, patience, encouragement and support throughout my years as a doctoral student at theUniversity of British Columbia. I especially would like to thank my committee members ProfessorsGlen Donaldson, Tsur Somerville and Murray Carlson. I would also like to thank my co-authorand friend Tom Davidoff for the important role that he played in my time at UBC.I am indebted to my academic mentors who took me under their wings and brought me alongincluding, Chris Costello, Peter Berck, Dimitrios Tsomocos and Tan Wang. I owe a special thanksto Robert Dummer for going above and beyond.I want to thank a few of my classmates including John Dutchak, Richard Souma, Vikram Agnihotri,Marshall Duggs, Sebastian Pearce, Vincent Gregoire and Alejandra Medina. We have been througha lot together.Finally, I thank my family and my friends for their unconditional love and encouragement duringthis journey.ixChapter 1IntroductionThis thesis is a collection of three essays in Real Estate Finance. Although the topics are diversethey share in common the fact that they address fundamental questions related to mortgage contractterms, pricing and behaviour in residential and commercial property markets.The first essay studies the determinants of contract terms in commercial mortgages. A cornerstoneof finance theory is that risk and return should be positively related. However, existing empiricalstudies often find a negative relationship between interest rates and risk terms (e.g., loan maturity,loan-to-value ratio, etc) in mortgage contracts. Previous studies have found such results puzzling,and have surmised that they may arise because traditional models do not explicitly account for thesimultaneous determination of interest rates and mortgage terms. We therefore specify separatesupply and demand equations for loanable funds, estimate them simultaneously, and then examinemortgage contracts as equilibrium outcomes of a multidimensional negotiation between borrowerand lender. Our empirical results reveal that borrowers and lenders individually require higherreturns for greater risk, as theory requires, but that this can produce negative risk/return correlationsin contract outcomes, as observed in the data. We solve various puzzles, demonstrate how variousrisk factors impact the simultaneous determination of equilibrium interest rates and loan terms,and show that mortgage lenders and borrowers of different types have understandably different riskappetites.The second essay investigates strategic default by borrowers in the Home Equity Conversion Mort-gage (“HECM”). HECM’s are a complex financial product that provide borrowers with implicit1home price insurance. We argue that the HECM program invites adverse selection and moral haz-ard because the pricing of mortgage insurance is not risk based. We document geographic adverseselection in HECM originations by showing that a disproportionate share of HECM’s were origi-nated in markets that experienced both high appreciation and significant declines in price. We thenlook at the credit utilization of HECM borrowers and test to see if credit use changes as the valueof home changes.The third essay studies the effect on residential property prices arising from proximity to oilpipelines. This paper uses a variety of static hedonic and dynamic event study methodologiesto estimate the effect of pipeline proximity on residential property values. First, we present moredetailed and precise measures of proximity at a very localized level than are found in other papers.Second, we address the land use context of the pipeline easement itself and identify its criticalinfluence on the estimated effect of pipeline proximity on home value. Finally, we take advantageof two different kinds of shocks to awareness of the pipeline’s presence and the possible risks tosee whether increase in both affects prices. In doing so we highlight the sensitivity of hedonic pric-ing of environmental hazards to specification bias because of parametric treatment of proximityand omitted variable bias because these hazards are associated with land uses that themselves aredisamenities for residential properties.Because each essay investigates a different topic, the chapters were designed to be self-contained. Ithus leave a more exhaustive discussion of the research questions, literature review and contributionof each paper to the introduction specific to each chapter.2Chapter 2The Simultaneous Determination ofInterest Rates and Loan Terms:Evidence from the Mortgage Market12.1 IntroductionA cornerstone of financial economics is that risk and expected return should be positively related.One would therefore expect mortgage interest rates to be higher for properties with higher risk,ceteris paribus. Existing empirical studies have been unable to find an economically strong rela-tionship between the interest rate (spread) on commercial mortgages and popular risk indicators(e.g., loan-to-value ratio, maturity, etc.), however, and in some cases the correlation is even nega-tive, which is a puzzling finding. To quote the seminal work of Titman et al. (2005, p.712): “theloan-to-value (LTV) ratio of a mortgage is expected to be positively related to mortgage spreads,but our evidence on this is mixed. Similarly, we expect from theory that mortgage maturity shouldbe positively related to mortgage spreads, but we empirically find the opposite.”Similar results are obtained by many other authors - including Ambrose et al. (2016), An et al.(2011), Maris and Segal (2002); Nothaft and Freund (2003); Titman and Tsyplakov (2010) -all which report that coefficients on many risk indicators in spread regressions often have either the“wrong” sign or are at best weak economically and/or statistically.2 These findings are important,1This paper was co-authored with Glen Donaldson at the Sauder School of Business - glen.donaldson@sauder.ubc.ca2In addition, empirical studies of commercial mortgage default generally find no relationship, or a puzzling negativerelationship, between default rates and risk indicators; e.g., Ambrose and Sanders (2003), Seagraves and Wiley(2015), Archer et al. (2002), Ciochetti et al. (2002), An et al. (2011) and Grovenstein et al. (2005). There is also alarge literature on bond/loan terms, interest rates and outcomes more generally, including: Bradley and Roberts (2015),Deng et al. (2015), Smith and Warner (1979) and Dichev and Skinner (2002).3and widely cited in the literature, in part because they challenge the core concept of a risk/returntradeoff which underlies much of finance; they also hamper empirical efforts to uncover deepertradeoffs between various loan terms. The purpose of our paper is to solve the risk-return puzzleand to investigate how various risk factors impact the simultaneous determination of interest ratesand other loan terms in multidimensional mortgage contract negotiations.The launching point of our paper is to note that finance theory requires individual borrowers andlenders to each demand higher expected returns for bearing more risk when making decisions.In this way finance theory constrains the behavior of an individual borrower or individual lender.However, theory does not require that a mortgage contract, which is the outcome of a multidimen-sional negotiation between a borrow and lender in which interest rates and loan terms are jointlydetermined, should necessarily exhibit a positive correlation between interest rates and LTV , ma-turity, etc. It is possible that lenders give borrowers with excellent credit ratings “better” termsin every dimension - lower spreads and also higher LTV , longer maturity, etc. - thus producingthe negative correlation between interest rates and other terms in contractual outcomes. We arguethat finding such a negative correlation in contracts is therefore not necessarily a puzzling rejectionof standard finance theory; one can only reject finance theory by finding that individual borrow-ers and lenders each fail to trade off risk and return when negotiating contracts. We also arguethat, to properly understand mortgage terms, one must therefore investigate how various mortgageterms are viewed and valued at the individual borrower/lender level, not just how they are jointlypackaged in the final negotiated contract.With the foregoing in mind, recall that the traditional approach to testing risk-return relationshipsin mortgages is to regress the mortgage interest rate (spread) on various risk indicators (e.g., LTV ,loan maturity, etc.) to find that coefficients on such risk indicators have either the “wrong” signor are very weak.3 This traditional method is thus a quasi-reduced-form approach that studies the3See, for example, the many papers listed in the first two paragraphs and first two footnotes of this paper.4outcomes of mortgage negotiations; it does not reach the deeper risk-return tradeoff that individualborrowers and lenders each make when engaging in such negotiations. We argue that, to uncoverborrower demand and lender supply, and thus test the risk-return relationship and investigate thetradeoff between different risk mitigators (e.g., LTV , maturity, etc.), one has to first specify separatesupply and demand equations for loanable funds. One can then estimate the supply and demandcurves simultaneously to obtain mortgage terms as an endogenously-chosen equilibrium outcomeand thus study how mortgage terms change when supply/demand curves shift in response to variousrisk factors. This is what we do in this paper, to find several important new results and explainprevious puzzles.We are not the first authors to notice that the traditional spread regression approach is incomplete.Indeed, many previous authors have commented that regressing loan interest rates (spreads) onloan characteristics (e.g., LTV , maturity, etc.) falsely assumes that loan characteristics are setexogenously, while in reality such loan characteristics are determined endogenously along withthe loan interest rate, and that such model misspecification may therefore be producing puzzlingresults. To again quote Titman et al. (2005, p.712), “These violations of the theoretical predictionsare likely to be due to the endogenous choice of mortgage characteristics. Specifically, lenders arelikely to require mortgages with higher downpayments, that is lower LTV and shorter maturities,on properties that are likely to be riskier.” Ambrose et al. (2016 p.11) similarly note that: “LTVis determined endogenously through negotiations between the borrower and the lender. Ceterisparibus, a higher LTV results in a riskier loan and a higher spread. However, riskier borrowersare typically forced to make higher down payments, which reduces the LTV on risky loans. Thus,riskier borrowers may end up with loans that have lower LTVs than safe borrowers would obtain.”Similar comments regarding the endogeneity of loan terms can be found in many other studies5dealing with interest rates and loan characteristics.4Some attempts have been made in the literature to partially address the endogeneity concern. Forexample, Ambrose et al. (2016) use an IV/2SLS approach of first regressing LTV on various riskfactors and using estimated coefficients from that first-stage regression to produce a fitted value forLTV , and then in a second stage inserting this fitted value for LTV into the right-hand-side of aregression of mortgage interest rate spread on various risk indicators. Using this approach, in whichfitted LTV serves as an instrument for actual LTV , Ambrose et al. (2016) find a significant, butstill weak, relationship between spread and the LTV instrument. We believe Ambrose et al. (2016)are on the right track - trying to account for endogeneity - but they do not go far enough; using asingle equation of spread on the variable instrumenting for LTV still only studies the outcome ofa negotiation (i.e., the equilibrium mortgage contract) rather than trying to uncover the supply anddemand curves that determine that mortgage contract as an equilibrium outcome.In this paper we therefore take an important step further; we explicitly specify an equation forthe supply of loanable funds, and a separate equation for the demand for loanable funds, in astructural model of the mortgage market, and then we estimate these supply and demand curvessimultaneously and investigate the equilibria they produce in response to changes in various riskfactors. Our paper is the first study we are aware of to take this approach with the loan-levelmortgage data. This approach allows us to study the influence of various risk factors on fundssupplied by lenders and demanded by borrowers and to investigate how interest rates and loanterms are determined simultaneously as the result of negotiation between a borrower and lender.Our empirical results are strong and robust, and reveal that lenders do indeed demand significantlyhigher interest rates on risker loans, and that borrowers borrow significantly less when interest rates4For example, Ambrose and Sanders (2003), Archer et al. (2002) and Grovenstein et al. (2005) cite the Stiglitzand Weiss (1981) model of credit rationing to note that lenders may respond to higher risk borrowers by adjusting thenon-interest rate contract terms (e.g., loan-to-income ratio, amortization, maturity, etc.) in addition to increasing theloan interest rate. See Ambrose et al. (2003) regarding simultaneity in data viewed at the aggregated level, as opposedto the individual loan-level examined in this study.6rise and amortization and maturity shorten. In other words, we find a strong risk-return tradeoff,consistent with finance theory, where previous studies have not. Through a series of comparativestatic exercises, we explain previously puzzling results concerning mortgage interest rates, LTVand maturity, and uncover an important role for amortization.In our study we use loan-level data - including information on mortgage interest rate, LTV , ma-turity term, property type, etc. - from individual commercial mortgages that were subsequentlysecuritized into Commercial Mortgage Backed Securities (CMBS). We use data at the individualloan-level, rather than studying the pricing of aggregated CMBS, because in the loan-level datawe can observe the individual mortgages that borrowers and lenders negotiated, which helps usuncover factors driving the individual supply of, and demand for, mortgages and the simultaneousdetermination of loan terms. The construction of our data set is discussed Section 2 below.In Section 3 we show that the relationship between mortgage interest rates and each of loan-to-value ratio, loan-to-income ratio, maturity and amortization, in our data’s mortgage contracts, allhave the “wrong sign” (i.e., risk and return appear to be negatively correlated, rather than positivelycorrelated), thus confirming that the risk-return puzzle documented in previous papers is also foundin our data. We then analyze the relationship between the various risk-indicating variables to revealevidence of their simultaneous determination with each other and collectively with interest rates.The desire to treat this simultaneity appropriately leads us to specify a structural model of themortgage market in which interest rates and mortgage terms are determined in a simultaneous-equations setting.In Section 4, we specify and simultaneously estimate supply and demand curves for loanable funds.We find that the supply of funds curve slopes strongly and significantly upwards; i.e., lenders arewilling to supply more funds when interest rates rise. We also show that the demand for fundscurve slopes strongly and significantly downwards; i.e., borrowers desire to borrow more funds7when interest rates fall. Perhaps most interestingly, the impact of various risk measures on thesupply and demand for funds is also clearly revealed; lenders are willing to lend significantly lessto properties with higher risk, and borrowers desire to borrow significantly more when risk-relatedrestrictions, such as maturity term, are relaxed. In other words, results from our model are stronglyconsistent with the theory that risk and return are positively related in the supplies and demandsof borrowers and lenders. Previous papers in the literature have not been able to find this clearand strong evidence of a risk-return tradeoff because, as noted above, previous papers have usedwhat has loosely been termed as a reduced-form approach to study contract outcomes rather thanestimating the deeper individual supply/demand curves that produce contract outcomes.In Section 5 we conduct a variety of comparative static exercises on the inverse supply and demandcurves to investigate the equilibrium impact of risk factors on interest rates and loan terms. Herewe demonstrate that, as theory predicts, increases in riskiness - e.g., by decreasing property value,increasing loan maturity, etc. - shift the supply curve such that lenders demand higher interest ratesand offer smaller loans as risk shifts higher, as theory predicts. Similarly, the borrowers’ empir-ical demand curve shifts as theory predicts it should in response to various factors. When bothsupply and demand shift simultaneously in response to a common shock, however, the resultingequilibrium can sometimes produce the puzzling type of contrary results we often observe in thedata. The move from old equilibrium to new equilibrium sometimes produces a positive correlationbetween risk factor and interest rate, but can also sometimes produce negative risk/return correla-tions given the direction and relative magnitude of supply/demand shifts and the relative slopes ofthe supply/demand curves. The final outcome becomes even more complex when risk factors arecorrelated with each other. By going through each case we are able to solve the LTV and matu-rity puzzles and explain the impact of amortization term and other factors on mortgage contractoutcomes. We also investigate supply/demand effects of property type and mortgage originator toexplain why riskier property types typically pay higher interest rates and receive stricter mortgage8terms. Section 6 concludes.2.2 DataIn this study, we employ loan-level data from individual commercial mortgages that were originatedin Canada, by lenders from the USA and Canada, and subsequently securitized into CommercialMortgage Backed Securities (CMBS). As noted above, we use data on the individual loans insidethe CMBS, rather than data on the aggregated CMBS, because we are interested in the factors im-pacting decisions of the individual borrowers and lenders of mortgage funds. We use Canadian datain part because all these CMBS were rated by a single rating agency (Dominion Rating Agency),thus avoiding the potential for the sort of agency-shopping effects, on the selection of mortgagesto include in the CMBS, that have been previously documented by Stanton and Wallace (2012),Cohen and Manuszak (2013), and An et al. (2015). Furthermore, due to features of the Cana-dian tax code, the CMBS market was dominated by sophisticated investors with expertise in themarket, which helps to ensure that individual mortgages in the CMBS pools had been subjectedto robust underwriting processes thereby minimizing potential sources of confounding noise.5 To-gether these factors should allow us to more easily uncover risk-return tradeoffs in the individualmortgages that underlie the CMBS.Our data were extracted from CMBS securities prospectuses which are publicly available fromthe System for Electronic Document Analysis and Retrieval (SEDAR), which is a filing repositorymanaged by the Canadian Securities Administrators. Annex A of each CMBS prospectus provides5Prior to 2008, foreign investors paid a withholding tax on earnings from Canadian fixed income instruments, whichmade the after-tax yield on Canadian CMBS too low to attract interest from investors who were not intimately familiarwith the market (see Brown (2007) and Gamm and Kane (2013)). The CMBS were therefore of generally high quality- the delinquency rate of the Canadian CMBS market peaked at under a quarter percent (for comparison, the US CMBSmarket experienced cumulative losses of more than 2.87%, a loss rate more than 30 times the Canadian loss rate; seeWestlake (2007), DBR (2016) and Gamm and Kane (2013)).9detailed information for each individual mortgage inside that CMBS.6 This loan-level informationincludes fields such as: property type, net operating income, appraised property value, loan size,loan maturity term, amortization term, mortgage originator, and of course the mortgage interestrate.7 Also included are traditionally calculated underwriting ratios, such as: LTV (loan-to-valueratio), LTI (loan-to-income ratio), Capitalization Rate (income-to-value ratio) and the Debt ServiceCoverage Ratio at origination (which is a function of the loan size, loan interest rate, net operatingincome and amortization term).8The CMBS market in Canada became active in 1998; new CMBS stopped being formed in 2007when the CMBS new issue market dried up due to liquidity shortages around the time of the fi-nancial crisis.9 The market began to show some signs of rebirth after 2012 but, as of 2017, hasnot yet recovered to its former activity or broad participation. We therefore concentrate on theperiod 1998 - 2007 and thus downloaded data for all individual commercial mortgages backing allCMBS issued from January 1998 to December 2007. This gives us 2,666 individual commercialmortgages. From this initial set, we discarded loans that had missing or incomplete information;for example, in some cases the maturity term, property type or other fields were empty. We alsodiscarded loans which had internally inconsistent information - for example loans which reportedLTV that was not equal to loan size divided by property value - for fear that one or more piecesof information for that loan contained an error. This leaves us with the 2,383 individual mortgages6 See After downloading the PDF of each CMBS prospectus, loan leveldata were then manually extracted from Annex A for every individual property collateralizing each CMBS.7We use the “mortgage interest rate” as our interest rate measure; one could alternatively use the effective loan rate,but that would require information we do not have concerning points and fees charged at origination. The definition oforiginator includes the entity listed plus its related entities and affiliates.8The Debt Service Coverage Ratio is the ratio of net operating income to the annual debt service. Since commercialmortgages are amortizing loans, DSCR is the following non-linear function, in which L is loan size, N is amortizationand i is the mortgage rate:DSCR=IL∗ (1−(1+ i)−N)i9For background on the historical development and performance of the Canadian CMBS market see (Brown, 2007;Gamm and Kane, 2013; Westlake, 2007)10which form our sample. Table 1 reports summary statistics.Table 2.1: Summary StatisticsMean Std Dev p(1) p(25) Median p(75) p(99)Mortgage CharacteristicsLoan Interest Rate (%) 6.01 0.95 4.69 5.32 5.68 6.52 8.84Risk Free Rate (%) 4.23 0.65 3.29 3.90 4.09 4.39 6.10Spread (%) 1.54 0.51 0.64 1.17 1.46 1.84 3.10Loan Size ($M) 6.34 7.29 0.58 2.05 3.70 7.90 36.00Amortization Term (months) 290.37 51.49 132.00 240.00 300.00 300.00 360.00Maturity Term (months) 104.13 29.93 60.00 60.00 120.00 120.00 156.00Net Operating Income ($M) 0.80 0.98 0.07 0.26 0.47 0.94 4.90Appraised Market Value ($M) 9.73 12.61 0.95 3.14 5.57 11.37 63.20Property CharacteristicsLoan to Value Ratio (LTV) 0.67 0.10 0.33 0.62 0.69 0.74 0.80Loan to Income Ratio (LTI) 8.08 1.70 3.91 6.99 8.13 9.27 11.62Capitalization Rate (I/V) 0.08 0.01 0.05 0.07 0.08 0.09 0.13Value to Income Ratio (V/I) 12.12 2.07 7.67 10.72 11.96 13.41 17.69Debt Yield (ITL) 0.13 0.04 0.09 0.11 0.12 0.14 0.26Debt Service Coverage Ratio (DSCR) 1.61 0.32 1.24 1.42 1.54 1.70 2.76Property Type Obs. % Total Originator Obs. % TotalHospitality 71 2.98 Capmark 28 1.17Industrial 464 19.47 CIBC 83 3.48Mixed Use 177 7.43 Colliers 98 4.11Mobile Home 75 3.15 Credit Suisse 206 8.64Multifamily 332 13.93 First National 190 7.97Office 393 16.49 GE Canada 47 1.97Retail Anchored 286 12.00 GMAC 65 2.73Retail Unanchored 449 18.84 Laurentian 44 1.85Retirement 41 1.72 Merrill Lynch 660 27.70Self Storage 95 3.99 Royal Bank Canada 395 16.58Toronto Dominion 567 23.79Total Observations 2,383 100% Total Observations 2,383 100%From Table 1, we see that the mean loan interest rate is in our sample is 6.01%, with 98% ofobservations in a range of 4.69% to 8.84%. LTV ranges from a low of .33 to a high of .80, withmost of the LTV mass falling in a range between .62 and .74. LTI ranges from 3.91 to 11.62, witha mean of 8. The CapRate runs from .05 to .13, with most of the mass between .07 and .09. Themedian maturity term is 10 years (120 months) and the median amortization term is 25 years (300months), both of which are standard in this industry. Table 1 therefore reveals a healthy diversity11in loan terms - as well as diversity in property types and originator types - of the sort required toeffectively study tradeoffs between terms, while also confirming that our data are not wildly at oddswith data from other studies. We now turn to a preliminary analysis of the relationship betweenvarious loan terms in the mortgage data.2.3 The Risk-Return Puzzle and Endogeneity of Contract TermsWe begin our analysis by presenting a price/quantity graph of our data, in Figure 2.1 below.In Figure 2.1, the price (vertical axis) is the mortgage interest rate and the quantity (horizontal axis)is the quantity of funds borrowed/lent. To maintain comparability across mortgages, the quantityof funds is divided by the property’s net operating income.10 Each of the 2,383 grey dots in Figure2.1 represent one of our 2,383 mortgage contracts. The six colored dots in Figure 1 represent sixexamples of the wide range of mortgages in our data, as captured in Table 2.2. The colored crossesin Figure 2.1 are supply/demand curves produced by our simultaneous equations model developedin Section 4 below; the place where each pair of fitted supply and demand curves intersect in Figure2.1 indicates the fitted equilibrium interest rate/LTI point produced by our model as estimated onour data, all as explained below.We see from Figure 2.1 that, in each case, an upward-sloping supply curve intersects a downward-sloping demand curve to produce a model equilibrium point remarkably close to the data point ofcorresponding color. We note, by way of foreshadowing, that the factors which shift our supplyand demand curves to fit each of these points in Figure 2.1 will clearly explain the risk-returntradeoff of borrowers and lenders - which are consistent with finance theory - as borrowers/lendersnegotiate to produce each point as an equilibrium contract. All the colored crosses, and the factorswhich shift supply and demand to produce relationships between interest rates and risk indicators10 Property net operating income is a stable variable for the purposes of scaling, since our data consist of loansoriginated for the purposes of being securitized into CMBS and thus are collateralized by income- producing propertieswhere the underwriting reflects stabilized existing cashflow not proforma/expected cashflow.12(such as maturity and LTV), will be explained and discussed more fully in Section 5 below. At thisjuncture, to motivate our analytical approach, we briefly compare Loan 2 vs Loan 3 from Table2.2.We begin our analysis by presenting, in Figure 2.1 below, a standard price/quantity graph of ourdata.Figure 2.1: Loan Price vs Quantity from Mortgage Data345678910Loan Interest Rate2 3 4 5 6 7 8 9 10 11 12 13Loan to Income RatioLoan 1Loan 2Loan 3Loan 4Loan 5Loan 613Table 2.2: Information Regarding Specific Mortgages in Figure 2.1LoanIDLoanInterestRateInterestRateSpreadAmort-izationTermMaturityTermDebtServiceCoverageRatioLoan toIncomeRatioValue toIncomeRatioLoan toValueRatioPropertyType1 8.28 2.21 300 120 1.58 6.67 9.43 0.70 Industrial2 6.75 1.79 240 60 1.76 6.23 9.75 0.62 Office3 6.75 1.19 360 120 1.40 9.18 12.56 0.73 Multifamily4 5.57 1.91 180 60 1.66 6.12 10.49 0.58 Retail Unanchored5 5.44 1.62 300 60 1.61 8.50 11.33 0.75 Multifamily6 5.51 1.00 360 120 1.38 10.62 16.04 0.66 OfficeFrom Table 2.2 we see that Loan 3 has an interest rate spread of 1.19%, while Loan 2 has alarger spread of 1.79%, which implies that Loan 2 is riskier than Loan 3. One might thereforeexpect that Loan 2 would also have higher risk indicators, such as higher LTV , longer maturity,etc. However, from Table 2 we see the exact opposite. Loan 2’s LTV of 62.07% is lower thanLoan 3’s of LTV of 73.04%, thus revealing the puzzling negative relationship between spread andLTV cited in the literature. Furthermore, Loan 2’s maturity of 60 months is half Loan 3’s maturityof 120 months, also confirming the puzzling negative relationship between spread and maturitycited in the literature. In addition, Loan 2 has a shorter amortization than Loan 2 (240 monthsvs 360 months) and a lower LTI ratio (6.23 for Loan 2 vs 9.18 for Loan 3). All these variablestherefore consistently show that risk (as measured by LTV , LTI, Amortization and Maturity) goesin the “wrong” direction vs return (as measured by Spread). Instead of seeing a positive correlationbetween risk and return, in every comparison of Loan 2 vs Loan 3 we see a negative risk-returncorrelation. This is what the literature has found puzzling.Interestingly, Loan 3 was demanded by the owner of an apartment building (multifamily property),which tends to have the most stable cashflows of all property types, while Loan 2 is a mortgage onan office building which industry considers to be risker as discussed more fully below. This sug-gests a potential reason for the puzzle - the borrower with a safer apartment building has negotiated14a loan that is simultaneously better for the borrower in every dimension - lower interest rate, higherLTV , longer maturity, etc. The question we are most interested in answering is: what were therisk/return tradeoffs made by the demander and supplier (i.e., borrower and lender) of mortgagefunds to shift the supply and demand curves for funds to produce this collection of features in themortgage contract? This is one of the questions we will answer with our simultaneous-equationsmodel developed below (note that the supply/demand curve crosses in Figure 1 show that our modelproduces equilibria close to both Loans 2 and 3). Before turning to this, however, it is useful to fur-ther investigate the equilibrium relationship between loan terms because that will help us specifyour structural model and identify interesting comparative statics to be investigated with the aid ofthat model.2.3.1 Spread RegressionsBy comparing Loans 2 and 3 in Table 2.2, we saw the apparently puzzling negative relationshipbetween spread and loan terms across these two mortgages. To check that this negative relationshipis a general feature of our data, similar to what previous papers have found and not just an artifactof Loan 2 vs Loan 3, we run a basic version of the general sort of OLS regression of the mortgageinterest rate spread on various risk indicators that is often employed in the literature. Results usingall 2,383 mortgages in our data sample are reported in Table 3. Robust standard errors are inparentheses, with significance at the 10%, 5% and 1% levels indicated by one, two and three starsrespectively.The first column of Table 2.3 regresses Spread on LTV . Column 2 regresses Spread on LTI. Ascan be seen from comparing regression results from columns 1 and 2 in Table 2.3, LTI is moresignificantly correlated with Spread than is LTV; this observation will influence some modellingchoices below. Columns 3 and 4 of Table 2.3 add the Capitalization Rate, Amortization, Maturityand Riskfree Rate to the initial regressions. Columns 5 and 6 add Property Type and Originator15Type indicator variables.From Table 2.3, we see that we obtain as a general result the same apparently puzzling risk-returnfinding as previous papers. Key risk variables - such as LTV , LTI amd Maturity - all have the“wrong” signs; the return (Spread) seems to be negatively related to risk rather than positivelyrelated to risk - this is the puzzle. No matter which regression we use, we can never get thecoefficient on LTV , LTI or Maturity to become significantly positive.1111 By torturing the data hard enough, there are some suspect ways to get LTV positive although statistically insignifi-cant from zero - e.g., by cherry-picking time periods and omitting certain variables - but in no case were we ever able tomake LTV significantly positive in a quasi-reduced-form regression of the type shown in Table 2.3.16Table 2.3: Basic Regressions of Spread on Risk IndicatorsDependent Variable : Spread (1) (2) (3) (4) (5) (6)CharacteristicsLoan to Value Ratio× 100 –0.0044*** –0.0045*** –0.0044***( 0.00) ( 0.00) ( 0.00)Loan Size to Income Ratio –0.1183*** –0.0413*** –0.0388***( 0.01) ( 0.01) ( 0.01)Amortization Term 0.0001 0.0001 –0.0007*** –0.0006***( 0.00) ( 0.00) ( 0.00) ( 0.00)Maturity Term –0.0038*** –0.0038*** –0.0039*** –0.0040***( 0.00) ( 0.00) ( 0.00) ( 0.00)Capitalization Rate× 100 0.1386*** 0.1054*** 0.1254*** 0.0938***( 0.01) ( 0.01) ( 0.01) ( 0.01)Risk Free Rate 0.1443*** 0.1459*** 0.1197*** 0.1215***( 0.01) ( 0.01) ( 0.02) ( 0.02)Property Type IndicatorsHospitality –0.0740 –0.0546( 0.07) ( 0.06)Industrial –0.1149*** –0.1116***( 0.04) ( 0.04)Mobile Home –0.0535 –0.0483( 0.06) ( 0.06)Multifamily –0.0239 –0.0166( 0.04) ( 0.04)Office –0.0456 –0.0435( 0.04) ( 0.04)Retail Anchored –0.0198 –0.0157( 0.04) ( 0.04)Retail Unanchored –0.0161 –0.0138( 0.04) ( 0.04)Retirement 0.0032 –0.0010( 0.07) ( 0.07)Self Storage 0.1732*** 0.1807***( 0.06) ( 0.05)Originator Type IndicatorsCapmark –0.0786 –0.0956( 0.10) ( 0.10)CIBC –0.1740** –0.1638**( 0.08) ( 0.08)Colliers 0.0120 0.0220( 0.07) ( 0.07)Credit Suisse 0.0616 0.0659( 0.07) ( 0.07)First National 0.0649 0.0695( 0.07) ( 0.07)GMAC 0.0801 0.0786( 0.08) ( 0.08)Laurentian 0.2262** 0.2365***( 0.09) ( 0.09)Merrill Lynch 0.0924 0.0895( 0.06) ( 0.06)Royal Bank –0.0850 –0.0823( 0.07) ( 0.07)Toronto Dominion 0.1308** 0.1331**( 0.06) ( 0.06)Constant 1.8286*** 2.4924*** 0.4149*** 0.7102*** 0.8389*** 1.1064***( 0.07) ( 0.05) ( 0.10) ( 0.12) ( 0.13) ( 0.15)R2 0.0073 0.1576 0.2913 0.2939 0.3303 0.3324Observations 2,383 2,383 2,383 2,383 2,383 2,383* Omitted category for the indicator variables are: Originator Type = GE Canada, Property Type = Mixed Use17Also note from Table 2.3 that the coefficient on Amortization is very close to zero and economicallyinsignificant in all cases; when Amortization is statistically significant it is of the “wrong” sign.This also seems puzzling. Borrowers should be willing to pay more for longer amortizations,because as amortization increases ceteris paribus, debt payments decrease for any given loan sizethus implying a borrower can afford a larger loan for any given interest rate. Lenders also caredeeply about amortization because, as detailed above, it is a key element of the Debt ServiceCoverage Ratio, which measures the ability of a borrower to cover debt service payments. Ofcourse, amortization is being negotiated simultaneously with other terms, which may be obscuringdeeper tradeoffs among variables in the Spread regressions. We therefore further investigate suchsimultaneity among risk terms.2.3.2 Multidimensional Tradeoffs Among Risk VariablesBefore proceeding to development and testing of our structural model, it is useful to investigate therelationship among risk variables alluded to in the previous section and to uncover the key roles ofIncome and Amortization. We do this by graphing the multidimensional relationship between thequartet of risk variables: LTV , CapRate, Amortization and LTI. This exercise helps to reveal theimportance of LTI - a variable which is referred to in industry as the Inverse Debt Yield - and helpsto determine the specification of the supply and demand equations we will employ in our structuralmodel of the mortgage market, developed in the following section of this paper.18Figure 2.2: LTV , CapRate, Amortization and LTI05101520Loan to Income Ratio.04 .05 .06 .07 .08 .09 .1 .11 .12 .13 .14 .15Income to Value Ratio (Capitalization Rate)20yr Amortization 30yr Amortization OtherLTV = 50% LTV = 75% LTV = 80%Figure 2.2 captures the relationship between the quartet of: LTV , CapRate, Amortization and LTI.The vertical axis of Figure 2.2 is LTI; the horizontal axis is CapRate. Since LTV = LTI*CapRate,we can draw isocurves in Figure 2.2 that represent a constant LTV for combinations of LTI andCapRate. The dashed uppermost curve in the top right portion of Figure 2.2 is the isocurve repre-senting an LTV of 80%, which is an industry-standard maximum LTV for high quality borrowers.Figure 2.2 also plots isocurves for 75% LTV , an industry-standard maximum for medium-qualityborrowers. For the sake of comparison, a 50% LTV is also plotted. The dots in Figure 2.2 representmortgages in our data sample: red dots represent mortgages with a 30-year amortization, blue dotsare 20-year amortization, and light grey dots are the remainder of the mortgages in our sample,most which have a 25-year amortization.19From Figure 2.2 we see that mortgages with 30-year amortization (red dots) tend to have higher LTIand lower CapRate than other mortgages, and an LTV that in many cases reaches 80%. Conversely,mortgages with 20-year amortization (blue dots) rarely have LTV above 75%; they also tend tohave lower LTI and higher CapRates than mortgages with 30-year amortization. If it is true thatmortgage originators are willing to grant higher LTV mortgages to higher-quality borrowers, thenFigure 2.2 suggests that high-quality borrowers may also get longer amortizations and larger loansrelative to income.The findings of Figure 2.2 suggest a reason that all of the risk-related variables in Table 2.3’sSpread regressions have the “wrong” sign: it appears that lower-risk borrowers negotiate bettermortgage terms in every dimension, including lower spreads, larger loans, longer term and longeramortization. This gives rise to the apparently puzzling positive correlation between spreads andmortgage terms and poses a challenge for investigating the risk-return tradeoff, as well as tradeoffsbetween risk factors, since all the variables appear to move together and thus give no indication of atradeoff. The solution we propose is to specify and simultaneously estimate the underlying supplyand demand relationships that produce the mortgage contract as an multidimensional equilibriumoutcome, in order to observe the risk-return tradeoff being made by suppliers and demanders offunds as the equilibrium contracts are being negotiated. It is to this task that we now turn.2.4 A Simultaneous Equations Model of the Mortgage Market - Spec-ification and EstimationGiven our findings in the preceding sections of this paper, we formalize the economic structure ofthe mortgage market with the following simultaneous equations model.202.4.1 Model SpecificationEquation (2.1) represents the demand for loanable funds, Ld; (2.2) represents the supply of loanablefunds, Ls; and (2.3) is the market-clearing condition that the quantity of loanable funds demandedequals the quantity supplied. Explanations of these equations and definitions for all variables areprovided below.demand:LdI= α0+α1(i−α2r)+α3N+α4VI+α5M+15∑h=6αhPh+m∑k=16αkEk (2.1)supply:LsI= β0+β1(i−β2r)+β3N+β4VI+β5M+16∑h=6βhOh+n∑k=17βkUk (2.2)market clearing:LsI=LdI(2.3)The demand equation (2.1) represents borrowers of funds who enter the market with commercialproperties to mortgage (e.g., a shopping mall, or office tower, or residential housing complex, etc.).Each of these commercial properties generates “net operating income”, I, from its tenants such thatthe “quantity of loanable funds demanded”, Ld, relative to this income, is Ld/I.We employ the scaled variable, Ld/I, (i.e., the loan-to-income ratio, LTI, or Inverse Debt Yield) inour demand equation in order to maintain comparability across all loans. We use LTI here, ratherthan other potential candidates such as LTV, for three reasons. First, columns 1 and 2 of Table 3reveal that LTI explains more of the variation in interest rates than LTV explains. Second, fromFigure 2.2 we see that LTI does not have the truncation problem that LTV has, with a boundary at80% maximum. Third, we want to focus on property value - our Value variable - specifically asentering on the right-hand side of the supply curve as we consider the collateral desired by lendersand the factors that shift supply/demand curves.1212Although, as discussed in the text, LTI is a more appropriate dependent variable than LTV , we did neverthelessinvestigate a similar model with LTV , rather than LTI, on the left-hand-side and using I/V , rather than V/I, on theright-hand-side. We still find upward sloping supply and downward sloping demand, thus confirming our core results.However, model R2s drop significantly, and even become negative for some subperiods indicating model misspecifica-21From the right-hand-side of (2.1), we see that the first variable influencing the demand for loanablefunds is the spread of the “mortgage interest rate”, i, to be paid on the mortgage in excess ofthe “riskfree interest rate”, r. We can restrict the parameter α2 to be unity if we want to usethe raw spread, or we can permit α2 to take values other than unity to capture a scaled spread.Standard economic theory suggests that the α1 coefficient on the mortgage interest rate, i, shouldbe negative; in other words, the demand curve should slope downwards such that the size of loandemanded (i.e., the quantity) increases with a decrease in the interest rate paid (i.e., the price).Conversely, demand should be increasing in the riskfree rate r since, ceteris paribus, a higherriskfree rate implies a smaller interest rate spread for any given mortgage interest rate.From equation (2.1) we see that, in addition to the price terms, i and r, the borrower considers the“amortization” term, N, and the “maturity” term, M, of the mortgage. Previous studies have ar-gued that borrowers prefer longer amortizations because, as amortization increases, debt paymentsdecrease for any given loan size thus implying a borrower can afford a larger loan for any giveninterest rate. The coefficient onN should therefore be positive in the demand equation (2.1). It hasalso been argued that borrowers prefer longer maturity terms as a way of reducing rollover risk andfixed contracting costs, suggesting that the coefficient on M should also be positive in the demandequation (2.1). The borrower also considers the “value” of the property, V , which we divide byincome, I, to facilitate comparison across loans. The coefficient on V/I should be positive since aborrower should desire to borrow more as the value of its property increases, ceteris paribus.A borrower’s demand for funds can also be influenced by the type of property the borrower owns.For example, the literature has shown that apartment buildings generate more stable cash flowsthan hotels; vacationing at a hotel is a discretionary luxury item, compared to staying at home,tion, likely due to the truncation of LTV and lower correlation with interest rates noted in the text. Using LTV ratherthan LTI on the left-hand-side also has the undesirable consequence that, when inverting and graphing the supply/de-mand curves as we will do below, property value, V , appears both on the axis and as a shifting variable so it becomesproblematic to see the Value comparative static in which we are interested and which constitutes a new and importantresult of this paper.22and thus apartment-building income is subject to less business cycle risk.13 We would thereforeexpect the quantity of funds demanded to be higher for apartment buildings than for hotels, ceterisparibus. Such “property type” considerations are captured by the collection of indicator variables,P, in equation (2.1). Property types are numbered in ascending order of a priori riskiness, begin-ning with P1 = 1 if a “multifamily” property such as a residential apartment building, which aspreviously explained is the least risky property type. P2, P3, P4, P5, P6, P7 and P8 signify “mobilehome park”, “anchored retail”, “unanchored retail”, “industrial”, “retirement”, “office” and “selfstorage” respectively. P9 signifies “hospitality” (e.g., hotels), the riskiest category in terms of cash-flow stability and thus the category we expect to have the lowest funds demand, ceteris paribus.We follow tradition in the literature and designate the category which will be omitted from theregressions below as P0, a “mixed use” property.Finally, there are many factors, in addition to the factors discussed above, that could potentiallyimpact a borrower’s demand for funds but which we do not or cannot observe. For example,there is no variable in our data set that directly measures the skill and experience of the personrepresenting the borrower in contract negotiations; presumably more highly skilled negotiatorsobtain better deals. Even if we could measure the negotiator’s experience, we cannot measureany number of other less tangible factors about the borrower and its property. All these unknownfactors are represented by the “E” variables at the end of equation (2.1). If we knew what all theseadditional factors were, and could measure them all without error, we could in theory fit the dataperfectly. Since we do not know or measure such E factors, however, when we take this modelto data in the following section these unknown factors will be replaced by a regression residualterm.13The mortgage default literature has shown that commercial mortgage default varies systematically with collateralproperty type; see Ambrose and Sanders (2003), An (2007), An and Sanders (2010), Ciochetti et al. (2002), Vandellet al. (1993). Typically, multifamily loans are the least risky, followed by anchored retail shopping malls and officeproperty loans; hotel loans are viewed as the most risky of all commercial property collateral.23Now consider equation (2.2), which specifies loanable funds supplied to the market by mortgageoriginators. From (2.2) we see that, when offering the “quantity of funds supplied”, Ls, to a bor-rower with a property that generates “net operating income” I, the supplier of funds considers the“mortgage interest rate”, i, to be charged on the mortgage as well as the prevailing “riskfree rate”, r.Since standard theory suggests that lenders should supply more funds as the mortgage interest rateincreases, ceteris paribus, we would expect β1 to be positive, thus producing an upward-slopingsupply curve. Conversely, β2 should be negative since a rising riskfree rate narrows the spread,ceteris paribus.The supplier of loanable funds considers the borrower’s ability to service its debt and the collateralwhich secures the mortgage in the event of default. The ability to service debt is traditionally cap-tured by the Debt Service Coverage Ratio (DSCR). As discussed in Section 2.2 above, the DSCRis a function of loan size, income, interest rate and amortization; all these variables, except amor-tization, are already in the supply equation, including in the dependent variable Ls/I. IncludingDSCR in the supply curve directly is therefore problematic for model estimation, hampers supplycurve inversion and renders comparative statics challenging. We therefore add to the supply curvethe only variable in DSCR that is not already included in (2.2), which is “amortization”, N. Sincehigher values of DSCR indicate a greater ability to service debt, and since DSCR is increasing inN (since as N increase the size of mortgage payments declines, ceteris paribus), we expect thecoefficient on N to be positive in the supply equation (2.2).As has been noted by Titman et al. (2005) and others, risk averse lenders should want to lend lessfunds as loan maturity increases, since as maturity increases the value of the default put-optionincreases. The variable “maturity term”, M, is therefore included in the supply equation with theexpectation that its coefficient will have a negative sign.The “value” of the property, which is the collateral backing the mortgage, is denoted as V . As is the24case with the dependent variable, loan size, we divide V by the property’s net operating income, I,to produce the stationary variable V/I (i.e., the inverse CapRate) in equation (2.2). The coefficienton this Value variable should be positive, since more valuable properties provide more collateraland should thus receive larger loans all else, including income, held constant. By investigating - inSection 2.5 below - the impact of changing V/I on the simultaneous determination of equilibriuminterest rates and L/I, we will be able to solve the LTV puzzle.Finally, the way in which a supplier supplies funds could be influenced by attributes of the supplier,for example the strategies, constraints and policies of the institution, the skill and experience of theperson representing the supplier in negotiations, etc. Unfortunately, we don’t have data on any ofthese factors. The name of the institution that originated each mortgage in our sample is reportedon the prospectuses from which the data were extracted, however we are not aware of any theory orprior empirical evidence which links a financial institution by name to an expected LTI outcome inthe same way we were able to link a property type to an expected outcome above (e.g., hotels beingriskier than apartment buildings).14 Thus, the best we can do is include in equation (2.2) of ourmodel a collection of “originator type” indicator variables to capture the identity of each originator- which we designate O0 to O10 without an a priori ranking of any sort - and a set of unobserved“U” variables to represent everything else which, if known perfectly and completely, would allowus to fit the data perfectly. These “U” variables, which we do not observe, will be replaced by aregression residual term when we take the model to data in the following section.14There is some recent evidence in the banking literature which suggests that specific bank account managers whohave long-standing relationships with specific borrowers, and who thus may have more “soft information” about suchborrowers, may agree to different loan terms with those borrowers (e.g., Herpfer (2017), Agarwal et al. (2011),Bharath et al. (2011) and Ivashina and Scharfstein (2010)). Even this does not help us, however, since we don’tknow the identity of each loan officer. We did investigate a variety of ways to combine our originators into differentgroupings, for example: domestic vs foreign lenders, large vs small lenders, lenders who created their own CMBS vslenders who served as conduits into the CMBS of others, etc. Our core results are robust to all such perturbations.Indeed, all we need to preserve our core results is to have enough variation in the Okj variables to identify the demandcurve when estimating our simultaneous equations model.252.4.2 Empirical EstimationThe structural model of the mortgage market in equations (2.1) - (2.3) implies the empirical supplyand demand curves for loanable funds in equations (2.4) - (2.5). The subscript “j” indexes eachof our 2,383 mortgages, j and uj are the demand and supply residuals respectively, and wehave expanded the spread terms to isolate coefficients on i and r for ease of exposition (e.g., thesingle coefficient α2 in equation (2.4) replaces the product α1 α2 on r from expanding equation(2.1)).demand:(LdI)j= α0+α1ij+α2rj+α3Nj+α4(VI)j+α5Mj+15∑k=6αkPkj+j (2.4)supply:(LsI)j= β0+β1ij+β2rj+β3Nj+β4(VI)j+β5Mj+14∑k=6βkOkj+uj (2.5)Table 2.4 reports regression results from the simultaneous estimation of (2.4) - (2.5) by two-stageleast squares.15The left side of Table 2.4 reports results for the demand curve, the right side for the supply curve.Standard errors are in parentheses, with significance at the 10%, 5% and 1% levels indicated byone, two and three stars respectively. As can be seen from the bottom of Table 2.4, the R2 for thedemand and supply curves are .615 and .475 respectively, indicating a good fit based on the 2,39015For technical details regarding 2SLS estimation of simultaneous equations, see Greene (2012). Note that the Orig-inator Type indicator variables in the supply curve help identify the demand curve, since they appear only in the supplycurve, just as the Property Type variables in the demand curve help identify the supply curve (different identificationstrategies are investigated below to demonstrate robustness - see the footnotes in the sections below regarding Propertyand Originator type effects). Recall that Property Type = Mixed Use, and Originator = GE, do not have their ownindicators and are thus their effects are captured by the regression constant.26observations in our sample.16Table 2.4: Estimated Supply and Demand Curvesqdemand qsupplyCharacteristics CharacteristicsLoan Interest Rate -0.387*** 0.862** Loan Interest Rate(0.141) (0.412)Risk Free Rate 0.112 -1.107*** Risk Free Rate(0.144) (0.389)Amortization Term 0.013*** 0.016*** Amortization Term(0.000) (0.001)Value to Income Ratio 0.285*** 0.468*** Value to Income Ratio(0.023) (0.054)Maturity Term 0.002** -0.002** Maturity Term(0.001) (0.001)Property Type Indicators Originator Type IndicatorsHospitality -1.254*** -0.189 Capmark(0.161) (0.311)Industrial 0.202** 0.394 CIBC(0.096) (0.252)Mobile Home 0.343** 0.822*** Colliers(0.147) (0.227)Multifamily 0.643*** 0.192 Credit Suisse(0.101) (0.209)Office 0.097 0.408** First National(0.097) (0.204)Retail Anchored 0.254** -0.164 GMAC(0.102) (0.262)Retail Unanchored 0.238** 0.269 Laurentian(0.094) (0.307)Retirement 0.130 -0.009 Merrill Lynch(0.184) (0.201)Self Storage -0.061 0.448** Royal Bank(0.136) (0.209)-0.104 Toronto Dominion(0.209)Constant 2.198*** -2.742* Constant(0.579) (1.600)R-Squared 0.615 0.475Number of Observations 2,38316Recall that Property Type = Mixed Use, and Originator = GE, do not have their own indicators and thus their effectsare captured by the regression constant.27The first key point to notice about the regression results in Table 2.4 is that the supply and demandcurves have the correct slopes: supply slopes up and demand slopes down. To see this, note fromthe first row and first column of Table 2.4, we see that the coefficient on “Loan Interest Rate” issignificantly negative in the demand curve. This reveals that, as the loan interest rate rises thequantity of funds demanded (relative to income) declines. In other words, the demand curve forloanable funds slopes down as theory suggests it should. From the first row of the second column ofTable 2.4, we see that the supply curve has a significantly positive coefficient on Loan Interest Rate,which reveals that the supply curve for loanable funds slopes upwards as theory suggests it should.These findings regarding the slopes of supply and demand curves, as well our general results below,are robust to a variety of perturbations in model specification, sample start/end dates and variousdata filters.17 We will discuss these coefficients and slopes further below in our comparative staticexercises, during which we will also discuss in detail the coefficients on, and impacts of, all theother variables in Table Comparative Statics, Puzzle Resolutions, and the SimultaneousDetermination of Loan TermsFigure 2.3 plots the inverse supply and demand curves produced by parameter estimates from Table2.4’s regressions. The upward-sloping supply curve, and downward-sloping demand curve, passingthrough the “base-case point” represented by the black dot, trace out the supply and demand curvesand resulting equilibrium outcome (the black dot) assuming average values for V ,N, T and r, plusP0,O0. Each of the many light grey dots in Figure 2.3 represent one of the 2,383mortgages in our17For example, we investigated: (a) grouping property type and/or originators into various categories rather thanincluding them all individually as in Table 4, (b) outlier Winsorization and omission of mortgages with nonstandardmaturity terms or amortization terms, (c) subsamples from different time periods. Our core results are robust to all suchperturbations. We do find that the model becomes difficult to estimate if too many years are excluded, in some casesproducing negative R2 from the 2SLS estimation, which indicates misspecification. We also investigated adding yearand/or month dummy variables, but these were not helpful given that movements in the riskfree rate adequately captureintertemporal differences in the economic environment.28sample. This is the starting point for comparative static exercises.Figure 2.3: Inverse Supply and Demand Curves - Base Case456789Loan Interest Rate2 3 4 5 6 7 8 9 10 11 12 13Loan to Income RatioThrough a series of comparative static exercises, we will next discuss the coefficients on loancharacteristic variables in Table 2.4, explain the distribution of dots in Figure 2.3, investigate theeffects of various factors impacting supply and demand, and solve various puzzles. Note thatchanges to the independent variables in Table 2.4’s (e.g., V , N, r, etc.) shift the supply/demandcurves left/right (rather than up/down) since the loan-to-income ratio, the dependent variable inTable 2.4’s regressions, is on the horizontal axis in Figure 2.4’s inverse supply/demand curves. Ofcourse, the resulting equilibrium points will move up/down (as well as left/right) to explain theentire cloud of data, as we shall soon demonstrate.292.5.1 Changing the Riskfree RateWe consider first the effect on supply, demand and market equilibrium of changes in the riskfreerate. For any given loan interest rate, the spread between the loan interest rate and riskfree ratenarrows as the riskfree rate rises. The supply of loanable funds should therefore decrease as theriskfree rate rises thus decreasing the spread, ceteris paribus. Another way to think of this is to notethat a conduit lender - who borrows the funds necessary to originate a mortgage and then sells themortgage to a CMBS pool and repays the borrowed funds to be left with a profit from a spread andfees - should be very sensitive to an increase in the riskfree rate, which impacts the rate at whichthe conduit lender borrows. From the first row of Table 2.4, we see this exact result: the coefficienton Risk Free Rate in the supply curve is significantly negative as expected.Likewise, the coefficient on Risk Free Rate in the demand curve is positive as expected. Borrowersdemand more funds when the spread between the loan interest rate and riskfree rate declines.This demand effect is not as strong as the supply effect, however (i.e., the magnitude of the RiskFree Rate coefficient, relative to the Loan Interest Rate coefficient, is much smaller in demandthan in supply), which makes sense since a borrower’s mortgage payments are only indirectlyimpacted by the riskfree rate. This realization provides another good reason, in addition to thoseenumerated above, to allow the coefficient on the riskfree rate to be estimated separately from theloan interest rate coefficient rather than constraining these coefficients to be of equal magnitudeand opposite sign, which is effectively what is done when using “spread” as a single combinedvariable in empirical investigations.Having established, from the coefficients in Table 2.4, that the riskfree rate has the expected impacton both supply and demand individually, we can now investigate the impact of changing the riskfreerate on the market equilibrium. In a graph of loan interest rate vs quantity of loaned funds, thisriskfree rate effect manifests as a shifting of the inverse supply and demand curves, as shown inFigure 2.4 and discussed below.30Figure 2.4 investigates riskfree rate effects by plotting the baseline supply/demand curves fromFigure 2.3 as the black lines in the middle of Figure 2.4. Figure 2.4’s red long-dashed lines abovethis base case shows the effects on supply and demand of increasing the riskfree rate by 2 standarddeviations; the blue short-dashed lines below the base case in Figure 2.4 decrease the riskfree rateby 2 standard deviations.18 From Figure 2.4 we see that, as the riskfree rate falls from the +2 st devred equilibrium down to the -2st dev blue equilibrium, the supply curve shifts out and demand curveshift in - i.e., the curves shift from the red lines to the blue lines. This produces a strong downwardmovement in the equilibrium loan interest rate (as measured on the vertical axis of Figure 2.4).This result is exactly what we expected based on finance theory: a positive correlation between theriskfree rate and the equilibrium loan interest rate.Figure 2.4: Comparative Static - Changing the Riskfree Rate456789Loan Interest Rate6 7 8 9 10 11 12 13Loan to Income RatioBase Level (mean)Base Level - 2 Std. Dev.Base Level + 2 Std. Dev.Risk Free Rate18The significance (from zero) of the point estimates used as the basis of such 2 st dev shifts are reported in theregression results in Table 2.4.31The equilibrium effect on the quantity of funds borrowed/lent (measured on the horizontal axisof Figure 2.4) depends on the relative elasticities of the supply and demand curves and on themagnitude of the curve-shifts as the riskfree rate changes. From Table 2.4 we see that magnitudeof the coefficient on Risk Free Rate is much larger in the supply curve than in the demand curve,which suggests the supply curve should shift more, and this is indeed what we see in Figure 2.4.Given that the inverse demand curve is steeper than the inverse supply curve, these shifts suggestthat equilibrium loan size should increase as the riskfree rate falls, and this is exactly what we seein Figure 2.4.Our finding that loan size increases as the riskfree rate falls makes intuitive sense since, as theriskfree rate falls, the equilibrium loan interest rate also falls, which reduces the size of mortgagepayments and thus increases the borrower’s ability to service a larger debt. Combined, the resultsof the regressions in Table 2.4, and the comparative static exercise in Figure 2.4, provide the firstof many steps in explaining why the grey dots in Figure 2.3 flow from the top left portion of thegraph down to the lower right portion of the graph; the riskfree rate was falling from 1999 to 2007,the time-period of our sample.2.5.2 Amortization EffectsNext we consider Amortization. The Amortization parameter estimates from Table 2.4 are sig-nificantly positive for both supply and demand, indicating that increasing amortization increasesboth the supply of, and demand for, loanable funds, ceteris paribus. This makes sense because anincrease in amortization decreases the size of mortgage payments, which implies the borrow canafford to borrow more, and the lender can lend more given the borrower’s ability to cover its debtservice payments given the income being generated by the property (and, of course, assuming thatall other risk factors are held constant as this occurs).32Figure 2.5: Changing Amortization456789Loan Interest Rate6 7 8 9 10 11 12 13Loan to Income RatioBase Level (mean)Base Level - 2 Std. Dev.Base Level + 2 Std. Dev.Amortization TermThe amortization effect can be seen graphically in Figure 2.5. As in Figure 2.4, the black lines inthe middle of Figure 2.5 are the base case supply and demand curves. The blue short-dashed linesrepresent a 2-standard-deviation decrease in amortization term and the red long-dashed lines rep-resent 2-standard-deviation increase in amortization term. We see that as amortization increases,both the supply and demand curves shift out (the outward shifts being driven by the positive coeffi-cients on Amortization in Table 2.4’s supply and demand curves). In the resulting new equilibrium,the loan size increases by a large amount. From Figures 2.4 and 2.5 we therefore see that, whilechanging the riskfree rate shifts equilibrium points up and down, changing amortization shifts ourequilibria left and right.Increasing amortization also slightly reduces the equilibrium loan interest in our model, as can beseen in Figure 2.5. To understand why, note that the coefficient on the Amortization variable in33Table 2.4 is slightly larger in the supply curve than the demand curve, implying that the supplycurve shifts slightly more than the demand curve shifts as amortization increases. Also note fromTable 2.4 that the coefficient on Loan Interest Rate is of larger magnitude in supply than demand,indicating that the inverse supply curve is flatter than the inverse demand, which means that asupply shift in Figure 2.5 has a more powerful effect on Loan Interest Rate than a demand shift ofthe same magnitude. Combined, these two considerations (slope of the curves and magnitude ofthe shifts) produce a slight negative correlation between amortization and mortgage interest ratesin Figure 2.5.Our finding of a positive correlation between amortization and loan size in Figure 2.5 makes logicalsense because an increase in amortization reduces the size of debt service payments which impliesthe borrower can afford to borrow more, and thus the lender is comfortable lending more, all elseequal. The accompanying impact on loan interest rate is small, as understood in a simultaneousequation setting. This may explain why previous studies (and our own preliminary analysis inSection 3 above), based on regressions of interest rate on amortization and other factors, havefound either no relationship, or the supposedly “wrong” relationship, between interest rates andamortization, suggesting a potential puzzle in the role of amortization in the mortgage process. Asseen in Figure 2.5, we find that amortization is very important in understanding mortgages and thatthe weakly negative correlation between interest rates and amortization is not puzzling when theoutcomes are understood as equilibrium intersections of supply and demand curves.2.5.3 Solving the Loan-to-Value PuzzleNext we consider the impact of property value and provide a solution to the LTV puzzle.From Table 2.4 we see that property value is significantly positive in the supply curve; the esti-mated coefficient on the Value-to-Income Ratio is 47% and highly significant. This reveals that34each dollar of increase in property value, V , is associated with a 47-cent increase in the quantity offunds that suppliers are willing to supply for any given level of income, ceteris paribus. This valueimpact is therefore not only statistically significant; it is also very significant economically. Simi-larly, increasing property value also increases the demand for loanable funds, although the demandcoefficient of 28% is smaller than in the supply curve. The result that the supply impact is largerthan demand makes sense given that the mortgage lender, who bears the credit risk, probably caresmore about the collateral value of the property being mortgaged than does the borrower.The foregoing means that: (a) an increase in property value increases both the supply of, anddemand for, loanable funds - both the supply and demand curves shift out; (b) supply respondsmore to a change in property value than demand does; (c) in both supply and demand the responsein loan quantity is smaller than the change in value (since both coefficients are less than 100%);and (d) recalling from the previous section that the inverse supply curve is flatter than the inversedemand curve, a supply shift has a more powerful effect on the loan interest rate than does a demandshift of the same magnitude. Together, these four facts explain the LTV puzzle, as shall be outlinedbelow.35Figure 2.6: Changing Value-to-Income456789Loan Interest Rate6 7 8 9 10 11 12 13Loan to Income RatioBase Level (mean)Base Level - 2 Std. Dev.Base Level + 2 Std. Dev.Value to Income RatioThe impact of changing property Value can be seen visually in the shifting inverse supply anddemand curves in Figure 2.6. As in previous figures, the black lines in Figure 2.6 represent thebase case in which all variables are at their average values. In this base case the value-to-incomeratio is 12 (see Table 2.1, note the mean value-to-income ratio is 12, this is the initial value-to-income used to produce the black lines in Figure 2.6). The blue short-dashed line in Figure 2.6shows a 2-standard-deviation decrease in the value-to-income ratio, which reduces the value-to-income ratio to 8. The red long-dashed line in Figure 2.6 shows a 2-standard-deviation increase inthe value-to-income ratio, which raises the value-to-income ratio up to 16. In other words, fromthe blue equilibrium in the upper left of Figure 7 to the red equilibrium in lower right of Figure 2.6,the value-to-income ratio doubles from 8 to 16.From the horizontal axis of Figure 2.6, we see that, as the value-to-income ratio increases 100%36from 8 to 16 (represented by the curve shifts), the loan-to-income ratio, measured on the horizontalaxis, rises from roughly 6.5 to 9.5, which is approximately a 50% increase. In other words, a 100%increase in Value produces a 50% increase in Loan quantity, for any given level of income. Thus,from the blue equilibrium point in the upper left of Figure 2.6 to the red equilibrium point in thelower right of Figure 2.6, the loan-to-value ratio, LTV , has decreased. In other words, an increasein property Value is associated with a decrease in LTV . Figure 2.6 also reveals that the increase inValue has resulted in a decrease in the equilibrium interest rate, as measured on the vertical axis ofFigure 2.6. Thus, LTV and Loan Interest Rate are positively correlated in equilibrium.From Figure 2.6 we therefore learn several things: (a) increasing property value increases thewillingness of lenders to lend more funds and decreases the interest rate they charge in equilibrium,which causes borrowers to demand more funds which increases the equilibrium loan size; (b) theresulting increase in loan size is not as large as the increase in property value which caused it,which therefore decreases the equilibrium loan-to-value ratio thus producing a positive correlationbetween interest rate and loan-to-value ratio. Note that point (a) is entirely consistent with financetheory. For example, suppliers lend more and charge less as property value increases, thus revealingthe positive relationship between risk and return in lender behavior that finance theory requires. Wealso see a positive risk-return relationship in borrower behavior, as theory requires. The negativecorrelation between interest rate and LTV in the equilibrium contracts produced by lender/borrowernegotiation in our data is therefore entirely consistent with finance theory.The foregoing is an important point and one key contribution of this paper. The negative correlationobserved between interest rates and LTV in the data is NOT a puzzle, it is in fact produced as theproper outcome of supply and demand curve shifts in which both borrowers and lenders eachdisplay the risk-return tradeoff required by finance theory.We can go even further to note that finance theory has nothing definitive to say about the equilib-37rium relationship between interest rates and LTV in isolation. A positive, negative or zero corre-lation between equilibrium interest rates and LTV would not, on its own, be puzzling. All theoryabsolutely requires is that the supply curve slopes up and demand curve slopes down, that bothcurves shift out as value rises, and that the size of the loan increases with property value, ceterisparibus; the impact on equilibrium interest rates can be positive, negative or zero, depending onthe relative slopes of the supply/demand curves and magnitude of the shifts, all as estimated onwhatever data set is employed.The supply/demand curves estimated from our data produce a negatively correlated LTV/interestrate. In other data sets the correlation could be positive or zero. This explains the finding fromTitman et al. (2005, p.712) quoted at the beginning of our paper, in which Titman et al. notethat their “evidence on this [the interest rate/LTV correlation] is mixed”. We have shown thatby estimating supply and demand simultaneously to uncover borrower/lender behavior directly,rather relying on a quasi-reduced-form regressions of contract outcomes, the LTV puzzle is solved.2.5.4 Solving the Maturity PuzzleWe can explain the maturity puzzle in a similar fashion. From Table 2.4 we see that the Maturitycoefficient is positive in the demand curve and negative in the supply curve, with both coefficientsof equal magnitude. An increase in maturity is thus associated with outward shift in the demandcurve for loanable funds and an inward shift of the same magnitude in the supply curve. Sincethe inverse demand curve is steeper than the inverse supply curve, the outcome is an increase inloan quantity as the interest rate rises. All this is observed in Figure 2.7, which shows the effectof increasing maturity term from minus 2 standard deviations to plus 2 standard deviations aroundthe mean, ceteris paribus; Maturity and interest rates are positively related.38Figure 2.7: Changing Maturity Term456789Loan Interest Rate6 7 8 9 10 11 12 13Loan to Income RatioBase Level (mean)Base Level - 2 Std. Dev.Base Level + 2 Std. Dev.Maturity Term Recall now the statement from Titman et al. (2005, p.712) quoted at the beginning of our paper,in which they reported that “we expect from theory that mortgage maturity should be positivelyrelated to mortgage spreads, but we empirically find the opposite.” Thus far we have solved thefirst half of this puzzle by showing in Figure 2.7 that, ceteris paribus, mortgage maturity andinterest rates are indeed positively related, as theory predicts, in the equilibria produced by oursimultaneous equations model from Table 2.4. The next step is to explain why in our Table 2.3above, as in Titman et al. (2005) and many other papers, one observes a negative relationshipbetween maturity and interest rates in quasi-reduced-form regressions.The key to solving this puzzle is to uncover the importance of the ceteris paribus constraint thatallowed us to uncover the positive relationship between maturity and interest rate in Figure 2.7.When shifting our curves in Figure 2.7, we perturbed Maturity but held all other variables constant.39In reality, however, maturity is positively correlated with other variables - such as Value-to-Incomeand Amortization - and they often move together. This means that maturity frequently shifts inconcert with other variables that have a similar impact on loan size but an opposite effect on loaninterest rate.For example, recall from both Figures 2.5 and 2.6 that increasing amortization and value bothincrease loan size and decreases interest rate. Also note that the effect of an amortization shift, andof a value shift, on interest rate are each both larger than the interest rate effect from a maturityshift. Thus, when maturity shifts with either amortization or value - and especially if all threeshift together - the result is a net reduction in the equilibrium interest rate (and a large increasein loan size) which makes it appear as if maturity and interest rate are negatively correlated whenat a deeper level they are not. This net combined effect is what we observe in Table 2.3’s quasi-reduced-form regression (the sign on the Maturity coefficient is negative in the spread regression)and what previous studies have also observed and considered puzzling.Our analysis reveals there is no puzzle when maturity effects are isolated and measured at thedeepest level in a simultaneous equations model; there just appears to be a puzzle when consideringvariables at the level of their final contractual outcomes. This solves the maturity puzzle. It alsosuggests that to truly understand the data we need to study both the final contractual outcomes andthe underlying supply/demand curves that produce such contracts as equilibrium outcomes. Wewill therefore investigate simultaneous-factor curve shifts below, after we discuss property typeand originator effects.402.5.5 Property TypeWe next consider the Property Type variables in the demand curve. Recall from our discussionin the model specification section above, that the literature has shown that properties of differenttypes generate cashflows of different stability and thus are considered to have different levels ofrisk. Multifamily properties (e.g., apartment buildings) generate the most stable cashflows and arethus the least risky in this dimension. Hospitality properties (e.g., hotels) are at the opposite endof the spectrum, and other property types lie between with riskiness ranked in the order discussedabove in the model section.With the foregoing in mind, note in Table 2.4’s demand curve that the coefficient on Multifamily(the safest property type) is relatively large and significantly positive. This confirms that, as ex-pected, owners of low-risk apartment buildings demand loans with higher loan-to-income ratios,ceteris paribus. Conversely, the significantly negative coefficient on Hospitality confirms that, asexpected, owners of high-risk hotels demand smaller loan-to-income loans, ceteris paribus. Allother property types fall between these two extremes and are ranked according to cashflow riskas expected and discussed above. For example, Unanchored Retail properties are generally riskierthan Anchored Retail properties, which explains why the coefficient on Unanchored Retail in Table2.4 is smaller than the coefficient on Anchored Retail; Office buildings and Self-Storage complexesare more risky than retail properties and indeed they have progressively smaller coefficients, indi-cating that they demand loans with progressively smaller loan-to-income ratios, ceteris paribus,as expected. In other words, all the property type variables in the demand curve behave as ex-pected.41Figure 2.8: Changing Property Type456789Loan Interest Rate6 7 8 9 10 11 12 13Loan to Income RatioBase Level : Mixed UseHospitalityMobile HomeMultifamilyOfficeRetail AnchoredSelf StorageProperty TypeAs important, the property type variables in the demand curve provide more than enough variabilityacross mortgages to identify the supply curve.19 This can be seen in Figure 2.8, which plots demandcurves for various property types against the base-case supply curve (note that the supply curvedoes not shift since the property type variables are only in the demand curve).2.5.6 Originator TypeNext consider the Originator variables. Since the name of the originator is not a major concernfor the borrow (over-and-above the terms of the mortgage, which are already captured separately)19In the interest of robustness, we investigated the impact on model identification of dropping characteristics from thesupply/demand curves - e.g., eliminating Maturity from demand and V/I from supply - to see if identification or otherresults changed in important ways. Such perturbations did not change the finding that the supply curve slopes up anddemand down, but did lower R2’s substantially.42the Originator variables are in the supply curve, and only the supply curve, and thus facilitateidentification of the demand curve. The results of Table 2.4 indicated that the model has clearlysucceeded in this dimension, insofar as there is more than enough variability in Originator variablesacross mortgages to clearly identify the demand curve.Unfortunately, as discussed above in the section on model specification, there is no theory or priorempirical work to suggest which, if any, of the originator indicator variables should be similar to,or different than, the others. In the absence of such guidance, we checked for potential logicalgroupings among originators. For example, we checked whether large international banks havesimilar coefficients to each other, that other institutions are similar to each other, and that these twogroups are different from each other - but as can be seen from Table 2.4 such is not the case. Fur-thermore, Table 2.4 reveals that banks in general do not appear consistently different than nonbanklenders, nor do domestic lenders appear consistently different than foreign-based lenders. We eventried grouping institutions into those that issue their own CMBS vs institutions that serve as conduitlenders into the CMBS of other institutions, but again found no consistency within or between suchgroupings.Table 2.5: Property Types by OriginatorProperty Type→Originator ↓Hosp-italityIndust-rialMixedUseMobileHomeMulti-family OfficeRetailAnchRetailUnanchRetire-mentSelfStorTotal(%)Average 2.98 19.47 7.43 3.15 13.93 16.49 12.00 18.84 1.72 3.99 100Capmark 0.00 14.29 3.57 0.00 0.00 57.14 14.29 7.14 0.00 3.57 100CIBC 0.00 25.30 4.82 1.20 8.43 19.28 22.89 18.07 0.00 0.00 100Colliers 1.02 28.57 6.12 1.02 15.31 21.43 11.22 11.22 1.02 3.06 100Credit Suisse 4.85 8.25 10.68 4.85 11.65 15.05 5.83 26.21 10.19 2.43 100First National 1.05 12.11 16.84 0.00 19.47 14.74 0.53 28.42 2.11 4.74 100GE Canada 8.51 10.64 6.38 0.00 6.38 42.55 21.28 4.26 0.00 0.00 100GMAC 0.00 12.31 7.69 0.00 1.54 38.46 18.46 13.85 0.00 7.69 100Laurentian 4.55 18.18 4.55 0.00 20.45 18.18 0.00 34.09 0.00 0.00 100Merrill Lynch 2.88 12.88 1.52 7.42 18.03 11.52 26.21 8.33 1.52 9.70 100Royal Bank 3.29 29.62 8.10 0.76 13.92 17.47 11.14 13.67 0.76 1.27 100Toronto Dominion 3.53 26.10 10.58 1.94 10.93 14.64 0.00 31.39 0.35 0.53 100Another possibility is that different originators specialize in loans to different property types, for43reasons unrelated to the lender’s aforementioned attributes (e.g., unrelated to the lender’s size, in-ternational, nonbank, etc.) and which are thus non-detectable in the aforementioned groupings.This possibility is investigated in Table 2.5, which reports the percentage of each originator’s mort-gages that fall into various property types (note that the average percentages, reported in the toprow of Table 2.5, are averages for the full set of originators combined into one entity, which thusgives more weight to institutions that originated a larger number of loans, so the average in the toprow of Table 2.5 is not an average of the percentages down each column).To begin with an example, we can look across the First National row and down the Multifamilycolumn of Table 2.5, to see that 19.47% of the mortgages First National Financial originated wereon multifamily properties (the supposed least risky property type) compared to an industry average13.93%; conversely, looking down the Hospitality column we see that only 1.05% of the mortgagesFirst National Financial originated were on hospitality properties compared to an industry average2.98%. So, from looking at just these two cells, it appears First National may have been relativelymore active in less risky sectors, which may help explain why the coefficient on its Originatorvariable is positive in Table 2.4. Conversely, GMAC and Capmark have lower relative percent-ages in Multifamily and higher percentages in moderately risky Office and Self Storage properties,which may help to partially explain - subject to the caveats below - why GMAC and Capmark havenegative Originator coefficients in Table 2.4.While this separating method may in some cases be suggestive, it is not particularly strong or gen-erally robust across either originators or property types. For example, RBC has a portfolio thatmatches the industry average fairly well across categories and yet has an originator coefficientlarger than average. Clearly, there are many factors (e.g., institutional constraints, strategies, poli-cies, etc.) that help influence originator supply that are not reported in our data and thus are notincluded in our model. We therefore cannot draw meaningful conclusions regarding originators.The reason for whatever similarities or differences might be observed between originators, and the44importance and robustness of such differences, therefore remains a mystery that we cannot solvegiven the data set available to us in this paper. Fortunately, the primary reason we include theOriginator variables in the supply curve is to identify the demand curve, and in this most-importantdimension the model and data have been successful.2.5.7 Simultaneous Shifts from Multiple FactorsThus far we have investigated comparative statics considering one variable at a time. To see the fullpower of the model, we can put the pieces together to investigate the ability of various combinationsof variables and curve-shifts to explain the data. This leads us full-circle back to the informationfirst presented in Figure 2.1 and Table 2.2 near the beginning of this paper, which we can nowdiscuss in detail. Recall that the light-grey dots in Figure 1 plot our 2,383 mortgage contracts.Each brightly-colored dot in Figure 2.1 plots a specific contract as indicated in Table 2.2 (e.g., theblue dot in the top left corner of Figure 1 is “Loan 1” from Table 2.2).We can now finally explain what the colored crosses in Figure 2.1 represent - they represent thesupply and demand curves one obtains by plugging into the fitted equations in Table 2.4 the valuesfor the riskfree rate, amortization, maturity, value, income, and other variables, observed for eachmortgage reported in Table 2. The place where each pair of fitted supply and demand curvesintersect in Figure 2.1 indicates the fitted equilibrium interest rate/LTI point produced by our modelas estimated on the data. In each case, the supply/demand cross is close to the colored dot whichrepresents the actual data point from Table 2.2. From comparing each of the colored dots in Figure2.1 with the corresponding supply/demand intersection of the same color, we therefore see that ourmodel does a good job of explaining each of the data points from Table 2.2.The first and most immediately interesting observation from Figure 2.1 is that our model fits awide array of data points. For example, the same model that fits Loan 1, with a 2.21% Spread45and 70% LTV (see Table 2), also fits Loan 4 with a 1.91% Spread and 58% LTV, and Loan 6with a 1.00% Spread and 66% LTV. Note that Spread is lower and LTV is higher in Loan 6 thanin Loan 4, thus producing a negative correlation between Spread and LTV , but this correlation ispositive when comparing Loan 4 with Loan 1. This mixed correlation between LTV (risk) andSpread (return), which puzzled previous studies, is clearly explained in Figure 2.1 by the shiftingof supply and demand curves from one equilibrium point to another. The estimated supply anddemand curves each contain a strong and theoretically correct risk-return tradeoff so there is in factno puzzle.One feature which does seem to occur regularly across Table 2.2 is that supposedly safer propertiesseem to have higher LTV (i.e., larger loans). For example, we see from Table 2.2 that mortgageson Multifamily properties (i.e., Loans 3 and 5) have higher LTV than the mortgages on the otherproperty types.Interestingly, the mortgage with the most payment risk, as indicated by the debt service coverageratio, is Loan 6, with a relatively low DSCR of 1.38, indicating a relatively low level of operatingincome relative to the size of mortgage payment required to service the debt. It appears the sizeof the mortgage payment was made as small as possible by extending the amortization term out to360 months and still theDSCR is low. One might expect this higher risk to result in a high spread,but Loan 6 also has the smallest spread in Table 2.2. This might be explained by noting that Loan6 also has a relatively valuable property, resulting in a high value-to-income ratio and low LTV ,which are risk-reducing terms that together appear to offset the higher payment risk to producea smaller spread. As we observed in the various curve-shifting diagrams above, longer maturity,longer amortization, higher property value and lower riskfree rate, all shift the equilibrium pointdown and far out to the right, thus producing an equilibrium loan with a low interest rate and largeloan-to-income ratio; this is exactly what we see in Loan 6.46We could tell similar stories about every other point in Figure 2.1, with supply and demand curvesshifting in and out. In sum, from Figure 2.1 and Table 2.2, we see that in each case risk and returnare being traded off by the suppliers and demanders of mortgage funds as finance theory suggestsshould be the case. And, the equilibria produced translate the risk/return tradeoffs of the partiesnegotiating the mortgage into the observed package of simultaneously chosen multidimensionalmortgage terms and interest rates, all consistent with finance theory.2.6 ConclusionsFinance theory requires that individual borrowers and lenders each trade off risk and expected re-turn when making decisions. In this way finance theory constrains individual decision-making.However, it does not require that the terms a mortgage contract, which is the outcome of a mul-tidimensional negotiation between a borrow and lender in which interest rates and loan terms arejointly determined, should necessarily exhibit a negative correlation between interest rates and riskmitigators. Indeed, it is likely that riskier deals will exhibit higher interest rates and also higher risk-mitigating terms, including shorter maturities, lower LTV , etc., as lenders extend a “short leash” toriskier borrowers in every dimension. Conversely, safer borrowers can negotiate a “longer leash” inevery dimension. Finance theory does not preclude this. We therefore argue that it is not puzzlingto find an apparent lack of tradeoff when mortgage contracts are studied in a single spread-type re-gression at the contract level, even if there is a clear risk-return tradeoff at the individual borrowerand lender level.To find the theoretically-predicted risk-return tradeoff, one has to uncover the borrower demandand lender supply curves, where finance theory clearly requires that each individual agent shouldbe trading off risk and return in their decision-making. We have done this in our paper by spec-ifying and estimating a structural model of the mortgage market, in which we explicitly specifyan equation for the supply of loanable funds, and a separate equation for the demand for loanable47funds, and then estimate the supply and demand curves simultaneously. This has allowed us to un-cover the theoretically predicted upward-sloping supply curve of lenders, and downward-slopingdemand curve of borrowers, along which risk and return are being traded.We have also conducted a series of comparative static exercises to see the equilibrium effects ofchanging risk-indicating loan terms, including loan-to-value, amortization, maturity, etc. In everycase we have shown that borrowers and lenders react to risk, and set risk-adjusted interest rates,in the way theory predicts. Our work therefore solves several puzzles in the literature, includingpuzzles regarding the impacts of LTV and maturity term in commercial mortgages. Our resultsalso shed light on the way in which borrowers and lenders negotiate commercial mortgages, withsafer borrowers receiving in equilibrium lower interest rates and also larger loans, longer terms andlonger amortization. We believe our approach and results constitute new contributions to this areaof the finance literature.48Chapter 3Do Reverse Mortgage Borrowers UseCredit Ruthlessly?13.1 IntroductionThe US Federal Housing Administration (“FHA”) regulates and insures Home Equity ConversionMortgages (“HECMs”). HECMs are “reverse” mortgages that offer older US homeowners cash,lines of credit, or regular monthly payments. Borrowers may defer repayment until they move ordie, and at that time, their liability is limited to the lesser of the outstanding loan balance or theresale value of their home.2 In exchange for fees and interest charges, FHA compensates lenders ifthe mortgaged home is worth less than the outstanding balance at termination.HECM’s limited liability feature implicitly insures borrowers against decreases in their home’svalue. Most HECMs originated prior to 2008 were lines of credit that let borrowers flexibly drawcredit over time, up to a limit that grows at the loan interest rate. As shown by Davidoff (2012) anddetailed in Section 3.2, these HECMs effectively bundle (i) a line of credit that must be repaid, with(ii) an “exotic put option” on the borrower’s home. This option gives borrowers the right, but notthe obligation, to sell their home at the date of loan termination for an amount equal to the creditlimit. The put option is “in-the-money” if the credit limit is greater than the value of the home.If borrowers have warning before they move or die, and face no credit or psychological costs toborrowing more than they repay, then if the put option terminates in the money, it is optimal to1This paper was co-authored with Tom Davidoff at the Sauder School of Business - tom.davidoff@sauder.ubc.ca2We discuss the possibility of credit score damage below.49“exercise” the put by exhausting the credit line at or before loan termination. The payoff to holdingthe put option at termination is any positive difference between the credit limit and the resale valueof the home. The borrower’s gain in that case is a loss to FHA.Providing either the liquidity product or the put option embedded in HECM may serve legitimatepublic purposes. However, FHA appears to view its implicit sale of put options as a necessaryinconvenience in thickening the market for liquidity product (i). The HECM program’s statedintent is to help poor older US homeowners remain in their homes; the enabling legislation doesnot mention helping retirees smooth capital gains across states of the world as a program objective.3Home equity is a large share of wealth for many retirees and moving is unappealing to most (Bayerand Harper (2000)). Artle and Varaiya (1978) describe the complications to retirement financeand negative welfare effects of linking the spending of home equity to moving.4 HECM helps breakthe link and may thus complement enormous federal interventions into housing and retirementfinance. The target clientele lack resources other than their homes with which to repay debt, butFHA intends to at least break even on its HECM insurance.FHA’s HECM insurance program has suffered striking “adverse selection” on the dimension ofhome price appreciation. Despite intending to avoid losses, FHA did not vary insurance pricingformulas across time or locations during the home price boom. Originations were concentrated nearthe 2000s home price peak in markets that suffered severe price busts thereafter. Many borrowershave profited in the sense that they have credit limits greater than their homes are worth. 2014budget estimates indicate that FHA’s HECM insurance fund faces a gap in present value of roughly$5 billion between fee income and shortfall claim payouts on all HECMs originated to date.3HECM’s purpose is to soften “the economic hardship caused by the increasing costs of meeting health, housing,and subsistence needs at a time of reduced income,” per the National Housing Act. HECM was authorized in 1987under a Housing and Community Development Act. The benefits of expanding the market for home price insurance aredescribed in Shiller (1993) and Caplin et al. (1997).4Skinner (1996) and Davidoff (2009) observe that this link may undermine demand for life annuities and long-termcare insurance. Conventional mortgage finance requires a fixed loan term and income that can be pledged to repayment,both of which are unattractive for low income retirees.50Shan (2011) and Haurin et al. (2014) observe that this selection is consistent with Akerlof (1970)-style conscious “lemons” selling. Given prices did not reflect arguably clear differences in riskacross time and places, potential borrowers may have found HECM most appealing when andwhere they (correctly) believed the put value was greatest. Following common use in the study ofmortgages, and intending no moral judgement, we call the conscious calculation and exploitationof limited liability features “ruthless.”5 We use “adverse selection” to refer to the positive ex-postcorrelation between put option value and HECM take-up across time and regions, with or withoutconscious intent among borrowers. We note that HECM appears to have been “underpriced” duringthe home price boom (Davidoff (2012)), but this paper is not intended to test that notion.Ruthless selection into HECM would imply older US homeowners were able to recognize thehome price protection implicit in a complex mortgage product and to identify which housing mar-kets were most overheated near the recent cycle peak. Such financial sophistication would besurprising in light of other findings concerning American retirees and housing market participants.Reverse mortgages are a potentially crucial component of retirement finance, but they are difficultto market profitably if borrowers are willing and able to exploit the price and longevity insurancethey typically embed. It is thus worth exploring whether ruthless calculation seems to have drivenselection.The fact that most HECMs originated before the cycle peak were credit lines lets us ask whetherHECM borrowers use credit ruthlessly. In particular, we ask if borrowers whose loans terminatedwith credit limits greater than their homes’ resale value have been likelier to exhaust credit thanotherwise similar borrowers whose loans do not terminate with in-the-money put options. This isakin to asking whether consumers who adversely selected into an insurance product filed claimswhen the insured event occurred.5The expression has been used since at least Foster and Van Order (1984) and Vandell (1995). As we understandit, the term has no more pejorative content than “moral hazard” or “adverse selection.”51Ours is a one-way test of ruthless intent. Borrowers who do not try to exhaust credit before anin-the-money put option expires presumably did not select into HECM with the intent of borrow-ing more money than their home would be worth at termination. However, borrowers who seek toexhaust credit when faced with an in-the-money option at termination may not have performed so-phisticated option value analysis ex-ante; they may simply behave opportunistically ex-post. Eitherfinding should be interesting in the context of financial literacy among retirees and housing marketparticipants, and with respect to the optimal design and market growth prospects of products likeHECM.How well households understand mortgage and retirement finance should inform the extent towhich the government lets the free market operate in these areas, a matter of current debate.Lusardi et al. (2009), Chalmers and Reuter (2012), and Brown et al. (2008) document whatappears to be limited financial sophistication among retirees, particularly with respect to annuityvaluation. The valuation of life annuities is similar to, but simpler than, the valuation of embeddedHECM put options. Ruthless selection into HECM would represent a significant counterpoint tothese studies.Home price movements across US housing markets in the 2000s both make conscious adverse se-lection plausible and call into question homeowners’, lenders’, and even economists’ understand-ing of downside home price risk. The most severe price cycles occurred in historically elasticallysupplied markets in the “Sand States” of Arizona, California, Florida, and Nevada.6 The steepprice declines in the Sand States were arguably foreseeable, given long histories of rapid supplygrowth and low price appreciation prior to explosions in prices between 2004 and 2006. Footeet al. (2012) argue that most market participants, including professional investors, did not believe6Outside of Coastal California, where cycles were less violent and HECM originations grew less than in the CentralValley; see Davidoff (2013). There remains disagreement among economists on fundamental questions such as whetherhousing markets with more elastic supply were insulated from severe price declines. Contrasting views are presented inKrugman (2005), Glaeser et al. (2008), and Davidoff (2013).52that home prices were likely to fall substantially from peak levels around 2006. While HECMorigination growth through the peak was relatively large in the Sand States, the fraction of eligibleborrowers using HECM has been below 5% almost everywhere, and below 15% in all metropolitanareas. Most older Sand State homeowners thus missed what appears to have been an opportunityto borrow considerably more than their homes are likely ever to be worth again.Home equity lending is a potentially important source of retirement finance and borrowers’ finan-cial sophistication is double-edged with respect to market growth. While theoretically valuableto consumers and central to federal priorities, reverse mortgage markets appear to be small every-where, including the US. Bureau (2012) estimates that roughly 2.5% of those eligible participate inHECM.7 Reverse mortgages may be uncommon partly because they are financially complex. Red-foot et al. (2007) shows that the most important reason borrowers shy away from reverse mortgagesis that they are a “high cost” product. Davidoff (2012), however, shows that the costs and fees thatborrowers pay at loan origination may frequently be smaller than the actuarially fair value of theimplicit price insurance. If borrowers do not recognize the value of using HECM ruthlessly, theymay perceive fairly priced insurance as expensive.If HECM borrowers are financially sophisticated, the design and pricing of retirement home equityproducts becomes difficult. Lenders or guarantors are exposed to risks involving interest rates,home price movements, and borrowers’ mortality and mobility. Most of these risks may be im-pacted by adverse selection or moral hazard. Many studies have identified problems of informationand incentives in the closely related life annuity and conventional mortgage markets.8 Shiller and7A small private US loan market essentially disappeared after the late 2000s home price bust. Davidoff (2010) showsthat conventional home equity loans and cash-out refinances are also uncommon among retired homeowners.8For example, Finkelstein and Poterba (2004) document adverse selection on longevity in the life annuity market.Philipson and Becker (1998) even suggest moral hazard on longevity may operate. Foster and Van Order (1984)and Deng et al. (2000) are commonly cited studies of problems of moral hazard and dynamic adverse selection intomortgage prepayment. Mayer et al. (2012) discusses theoretical problems with prepayable subprime loans. Problemsof selection and moral hazard in mortgage originations among borrowers and intermediaries are explored by Keys et al.(2010) and others. Likely the most celebrated reverse mortgage (“viager,” the French version) borrower was JeanneCalment, who lived to over 120 at great cost to her lender and his heirs.53Weiss (2000) and Miceli and Sirmans (1994) observe that limited liability may encourage HECMborrowers to undermaintain their homes. Davidoff and Welke (2006) argue that HECM distortsborrowers’ incentives to remain in their homes. Contractual problems with negative amortizationhave been borne out in the current downturn: HECM terminations have slowed down considerablyfrom the peak, and market participants report that a surprisingly large number of borrowers havedefaulted on property tax or insurance obligations. Mortgaged homes appear to appreciate at a rateless than market averages, per Capone et al. (2010) and Davidoff (2014). Raising costs and feesand inhibiting choice might soften moral hazard, but could exacerbate adverse selection.9HECM is an appealing market in which to measure borrower ruthlessness in part because the onlyknowledge required of borrowers to exploit the embedded put option is whether the loan will termi-nate soon and whether the principal limit exceeds collateral value. Exercising the default and pre-payment option in conventional mortgages and student loans involves giving up future prepaymentand default options. Those options are thus difficult to value for consumers and econometricians:see Lucas and Moore (2007) Deng and Quigley (2004), Deng et al. (2000), Foster and Van Or-der (1984) and Vandell (1995). Mayer et al. (2011) find relatively strong evidence of strategicbehavior with respect to default options in a federal loan modification program.Section 3.2 provides background on HECM loan structure and the incentives to use credit. InSection 3.3, we lay out our approach to inferring borrower ruthlessness from differences in creditexhaustion between groups. We emphasize two challenges. First, optimal credit use is difficult tocharacterize beyond the desirability of using up all credit before loan termination if the put optionterminates in the money. We cannot interpret failure to exhaust an active credit line as absenceof ruthlessness and may thus wish to confine analysis to terminated loans, but borrowers whoseloans terminate may face unobservably different liquidity and put option incentives than borrowers9Any design features designed to limit losses when home values fall face marketing constraints related to a perception(the accuracy of which we do not address) that some lenders seek to exploit information asymmetries through hiddenfees and interest charges.54whose loans do not terminate in our data period. This is because borrowers who owe more thantheir homes are worth face very little capital cost to remaining in their home. In response to thisconcern, we provide evidence that endogenous terminations do not bias our results away fromfinding a difference in credit use. Ruthless borrowers may wish to exhaust credit early, so we alsocompare credit exhaustion between in-the-money and out-of-the-money loans among all borrowers,including those who do not terminate in our sample period.A second hurdle to identifying borrower ruthlessness is that demand for the put option (ii) may becorrelated with demand for liquidity product (i). This is important, because many borrowers appearto exhaust credit near loan origination for liquidity reasons. The metropolitan areas in which pricesfell far enough to generate high put option value saw large ex-ante price gains, which may have gen-erated high liquidity value, but their large subsequent price busts may have been associated withreduced liquidity demand. We thus emphasize comparisons among borrowers who experiencedsimilar price increases before the price bust. Table 3.3 provides some comfort that the borrow-ers experiencing relatively modest ex-post price declines are comparable on observable features tothose “treated” with in-the-money options. Importantly, differences in credit use between groupsare attenuated, not enhanced, by controlling for observable differences. We present evidence sug-gesting capital gains pressure after origination did not generate excess liquidity demand amongborrowers with out-of-the-money put options. We also find consistent results whether we compareborrowers terminating with in-the-money put options to out-of-the-money borrowers whose loanshave or have not terminated or to borrowers whose loans are still active but in-the-money throughthe end of our sample. Recognizing the strength of the assumptions required to use our data to in-fer the degree of ruthlessness among borrowers, particularly in light of the intensity of early credituse, Section 3.3.5 summarizes features of the data that provide some internal support for theseassumptions.Across specifications, we generally find that put option moneyness is not associated with the55propensity to exhaust credit. This suggests that borrowers did not select into HECM with theintention of borrowing more money than they expected their homes to be worth at termination.The Conclusion discusses policy implications and possible reconciliations of the simultaneous ex-istence of adverse selection and non-ruthless credit use.3.2 The Home Equity Conversion Mortgage (“HECM”)3.2.1 HECM loan structureUS homeowners over 62 are eligible for HECM loans as long as they can pay off any outstandingmortgage debt with available proceeds. For couples, the younger borrower’s age governs eligibilityand credit terms.10 Before loan origination, borrowers must complete programmed counsellingfrom an FHA-approved professional. Between program inception in the late 1980s and 2008, thelarge majority of borrowers took adjustable rate lines of credit, although other payment plans suchas regular monthly payments were available. Fixed rate loans that require all credit to be withdrawnimmediately, were introduced in 2008 and have proven popular since. A small private market grewthrough the home price cycle peak of the 2000s but has essentially died since then.HECM credit lines offer an initial principal limit that grows with borrower age and falls with the10-year LIBOR or US Treasury yield (the notch in realized put option value in 2006 shown inFigure 3.2 came from movements in 10-year interest rates). Among all HECM lines of credit inour sample, the median loan-to-value ratio at origination was 65%. Some loans hit a time- andmarket-varying cap on insurable value, such that loan-to-value ratios are smaller than available toborrowers with less valuable homes. This is particularly common in expensive markets, which runinto a global cap on insured value.10Anecdotally, there appear to be problems of older borrowers dropping younger borrowers from title to take on largerloans, then spending the money, dying, and leaving widowed partners with no home equity and a need to repay the loan.See “ A Risky Lifeline for the Elderly Is Costing Some Their Homes”, Jessica Silver-Greenberg Ackerman for New YorkTimes, October 14, 2012.56Each month, the outstanding balance on the credit line grows at the 1-year index rate plus a lender’smargin (typically 1.5% per year in our data), plus an FHA guarantee fee (.5% per year for loansoriginated through mid-2010). The maximum allowable outstanding balance grows at the samerate as the actual outstanding balance, so at any date up to loan termination, borrowers may drawon the credit line up to the point at which the balance is equal to the outstanding balance if allavailable credit had been drawn at loan origination.Loan servicers must respond within five days to credit requests from borrowers and may not reducethe credit limit in response to falling prices. HUD staff have indicated that there is no administrativeobstacle to exploiting a HECM as a put option by waiting until near termination to draw credit.Borrowers may prepay part or all of their HECM loans without penalty, but partial prepaymentis rare in practice. Starting in 2004, low-cost HECM refinancing became available,11 and a largenumber of borrowers exploited low interest rates and rising home prices to extract further homeequity.At origination, borrowers owe lenders closing costs and fees plus an FHA mortgage insurancepremium equal to 2% of a “maximum claim amount” during our sample period. The maximumclaim is equal to the lesser of the area cap on insured value or the appraised value of the home.Caplin (2002) suggests a total liability of 5% of property value at closing would not be out ofthe ordinary, but lender origination fees and closing costs vary considerably. Borrowers typicallyfinance closing costs by borrowing well in excess of them at origination.Borrowers need not make any payments until they vacate the home for more than six months,sell the home, or die. Couples can defer termination as long as one of the borrowers remainsalive in the home. At termination, borrowers owe the accumulated principal and interest balance.However, lenders have no recourse to any borrower assets other than the mortgaged home. Lenders11The FHA guarantee fee applied only to the increase in insured value, not the entire insured value.57must accept repayment in the amount of a third-party appraisal less reasonable selling costs if thatamount is less than the outstanding balance.Whether there are any adverse credit consequences of such a “short sale” if the loan terminateswhile at least one borrower is still alive is a question to which we suspect most borrowers donot know the answer at origination. In the conventional mortgage market, short sales typicallydamage borrower credit, even among non-recourse first liens. However, government marketing andfinancial protection documents regarding HECM emphasize limited liability and do not mentioncredit damage.12 Asked informally about such consequences, several mortgage lenders, FHA-approved HECM counselors, FHA and Fannie Mae staff, and a credit bureau executive gave usconflicting responses, with most indicating uncertainty and that they had never before been askedthe question. That a small sample of HECM counselors uniformly reported never having been askedabout the consequences of short sales casts some doubt on borrowers’ interest in using HECM forits put option value. If the loan terminates with the death of the borrower (or both borrowers if acouple), the estate owes no debt beyond the property value, and there is no possibility of damageto heirs’ credit.133.2.2 The embedded put optionDavidoff (2012) shows that the value of HECM lines of credit to ruthless borrowers can be de-composed into: (i) a line of credit that must be repaid, and (ii) a put option on the home that givesthe borrower the right to sell the home for the credit line’s principal limit. This ignores any credit12Frequently Asked Question 2 regarding HECM loans found in a search of FHA’s web site is as follows: “Can aborrower on a Home Equity Conversion Mortgage (HECM or reverse mortgage) ever owe more than the value of thehome? No, the borrower’s total debt on a Home Equity Conversion Mortgage (HECM or reverse mortgage) can neverexceed the value of the home. The borrower can never owe more than the home is worth. The HECM is a ‘non-recourse’loan. This means that the HECM borrower (or his or her estate) will never owe more than the loan balance or the value ofthe property, whichever is less; and no assets other than the home can be used to repay the debt.” Similarly, a ConsumerFinancial Protection Bureau HECM FAQ (dated February 2014) emphasizes limited liability and makes no mention ofcredit damage.13Indeed FHA has been pressed to require lenders to allow heirs to remain in the home and satisfy the debt by payingcurrent market value if that amount is less than the outstanding balance.58damage and presumes borrowers retain the capacity to respond to incentives and have sufficientwarning before termination to draw all remaining credit.To see the decomposition, suppose that the borrower uses no credit prior to termination. In thatcase, up to termination the only effect of the loan has been the need to pay closing costs, which welabel a fraction F of the home’s initial value h0. At the moment before termination, if the homeis worth more than the credit limit, there is no benefit to drawing credit. However, if the creditlimit exceeds the homes’ value, the borrower can draw the full credit limit and then pay the lenderthe home’s appraised value. The benefit of doing this is the positive difference between the creditlimit and the value of the home. This calculation is not different if borrowers intend to use homeequity for bequests or to fund long-term care, frequently cited reasons why HECM demand mightbe low.14If rt is the index yield, δ the borrower’s discount rate, s the interest rate spread, l the initial loan-to-value ratio, and hT the value at termination net of selling costs, then the borrower’s total risk-neutrally discounted home equity under this HECM strategy for a loan originated at date 0 andterminated at T is:V(T) = −Fh0+max(hT , lh0ΠTt=1 [1+ rt+ s])e−δT . (3.1)The second term on the right hand side of (3.1) is the greater of the credit limit or the home at thedate of termination. There is value to the credit line only if the former exceeds the latter, just asthe value of a put option at expiry is only positive if the strike price exceeds the underlying asset’svalue. The put option is in-the-money when the second term in parentheses exceeds the first.Credit used prior to termination does not affect the value of the put option, because early credit use14See Skinner (1996), Nakajima and Telyukova (2013). One might think that borrowers might believe prices willrise, and so perceive the home to be worth more than the credit limit. However, heirs with this belief can use the creditline to purchase a larger home.59is effectively repaid at the loan interest rate whether the put option is in- or out-of-the-money. Tosee this, suppose there is a draw in amount x on the line at some date A, but the line is otherwiseuntouched before T . The value to the borrower in (3.1) becomes:V(T) =−Fh0+xe−δA (3.2)+max(hT −xΠTt=A [1+ rt+ s] , lh0ΠTt=1 [1+ rt+ s]−xΠTt=A [1+ rt+ s])e−δTThe first term in parentheses is the payoff if the borrower sells the home and retires the HECM debt.The second term is remaining credit at termination. The borrower effectively repays principal andinterest on prior draws upon termination whichever of the two they receive. The benefit of the earlydraw is thus independent of the resale value hT and the value of the credit limit at T :∂V∂x=e−δA−ΠTt=A [1+ rt+ s]e−δT . (3.3)While put value is maximized by using any available credit just before termination, theory provideslittle guidance as to whether prescient borrowers whose loans have terminated in-the-money shouldhave used more or less credit relative to other borrowers well before termination. Using credit priorto termination can be part of an optimal strategy for two reasons. First, generalizing (3.3), we havethe standard Artle and Varaiya (1978) rationale for reverse mortgages. If a borrower discounts ata rate greater than the current loan index rate (δ > r), or wishes to smooth consumption betweenthe illiquid period before the home is sold and the more liquid (or dead) state after, then earlydraws are welfare enhancing. Moreover, consumption smoothing demand will likely generate acorrelation between optimal credit use and expected resale value of the home. Per Nakajima andTelyukova (2013), a reduction in the expectation of hT , which increases put value but reducesexpected home equity, should be associated with reduced consumption and hence liquidity demandthrough the loan’s life. However, the bad economic conditions associated with low prices could60generate family liquidity crises, so we regard the sign of the correlation between optimal credit usebefore termination and expected price declines as ambiguous.A second motivation for early credit use arises if borrowers believe they will die or lose the capacityto draw credit in response to incentives without warning. When such sudden incapacity may occur,the value decomposition no longer holds. The terminal payoff falls to hT minus accumulated debtwhen the put option is in the money, with a probability equal to the likelihood that the borrowerhas lost capacity for ruthlessness at T . In that case, early credit draws are not always effectivelyrepaid at the loan interest rate, because the remaining credit limit may not be drawn even whenoptimal.HECM borrowers almost never use credit in a way consistent with valuation (3.1). The number ofborrowers who draw less than 25% of available credit in the first year of the loan’s life and thenmore than 50% of remaining credit in any subsequent year is 8,411 out of 583,937 lines of creditin our sample. Of these, only 1,799 are in the Sand States where in-the-money options have beenconcentrated. This fact alone does not imply a lack of interest in the embedded put option amongborrowers, but it strongly suggests that adverse selection was not the product of borrowers with nointerest in consuming or investing home equity prior to sale exploiting put options. That pattern ofcredit use also seems inconsistent with fear of incapacity destroying put option value. Collateralvalue typically does not fall below the credit limit until more than a year after origination, sopreemptive put exercise is not easy to rationalize in the first year.15We conclude that:1. Ruthless borrowers will seek to exhaust credit at or before termination when credit limitsexceed collateral value.2. Put value maximization per se does not imply a correlation between the size of credit uselong before termination and put option value at termination.15Conceivably, borrowers might take large initial draws anticipating that incapacity to make sound decisions, but notdeath or exit from the home, is likely to occur prior to the put options arrival in the money.613. If death or incapacity can arrive with insufficient warning to withdraw available credit:(a) it may be optimal even for patient borrowers to withdraw the full credit line beforetermination, and,(b) borrowers who believe surprise termination occurs with probability smaller than thegap between the loan interest rate and their own discount factor, or whose homes arenot worth too much less than the allowable credit, may optimally fail to withdraw creditbefore termination with positive probability.3.3 Empirical Analysis of HECM Credit LinesWe begin our empirical analysis by documenting adverse selection into HECM. We then askwhether adverse selection was driven by intent to exploit the put options embedded in HECMloans. We use a sort of propensity score matching to construct a treatment group (borrowers whoseloans are in the money) and control group (otherwise similar borrowers who loans are not in themoney). We then estimate linear probability regression models to assess whether the treatmentgroup was more likely to exhaust credit to infer ruthless intent. A series of alternative comparisonsprovide results broadly consistent with our main results.3.3.1 Adverse Selection into HECM on Put Option ValueFigures 3.1 and 3.2 show that FHA has suffered severe adverse selection on put option value. Itis not plausible that the correlation between HECM value and market home price index declineswas driven by moral hazard. If HECM is an alternative to selling a home, HECMs should add to,rather than subtract from, market prices. Any effect would be small, as HECMs have small marketshare everywhere. The top panel of Figure 3.1 takes metropolitan area (m) × year-quarter (t)combinations as the unit of observation. The horizontal axis provides a measure of the magnitudeof the recent home price bust: the ratio of real Federal Housing Finance Agency (“FHFA”) repeatedhome price index for m at date t to the value for m in 2010, quarter 4. The vertical axis plots the62ratio of HECM loans originated at mt to US Census estimates of number of homeowners over 65inm as of 2010. Data run from 1992 through 2010. The unconditional correlation between HECMmarket penetration and ex-post price declines is .56.Figure 3.1: HECM origination vs metropolitan area price change by origination quarter●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ● ● ●● ●●● ●●● ● ●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●● ●●●● ●●●●●●● ●●●●●● ●●●●●●●●●●●●●●●●●●●●● ●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●●●●●●●●●● ●●●●●●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●●●●●●● ●●●●●●●●●●●●●●●●●●●●● ●●●●● ●● ●●●●●●●●●●● ●●●●●●●●●●●●●●●●●● ●●●●●●●●●●●●●●●●●● ●●●●● ●●●●●●●●●●●●●●●●●●● ●● ●● ●● ●●●●● ●●●●●●●●●●●● ●● ●●● ●●●●●●●●●●● ●●●●●●●●●●●●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●●●●●●●●●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●● ●●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●●●●●●●●●●●●●●● ●● ●●●●● ●●●● ●●●●● ●●●●●●●●●●●●●●●● ●●●●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●●●●●●●●●● ●●●●●●●●●●●●●●●● ●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●●●●●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ● ●●●●●●●●●●●●● ● ●●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●● ●● ●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●● ●●●●●●● ●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●●●●● ●●●●● ●● ●●●●●●●●●●●●●●●● ●●●●●●●●●●●●● ●●●●●●●●●●●●●●● ● ●●●●●● ● ●●●●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●●●●●●●●●●●●●● ●●●●●●●●●●●●●●●●● ●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●●●●●●●● ●●●●●●●●●●●●●●●●●●●● ● ●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●● ●●●● ●●●●●●●●●●● ●● ●●●●●●●●●●● ●●●●●●●●●●● ●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●●●●● ●●●● ●●●●●● ●●● ●●●●●●●● ● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ● ●●●●●●●●●●●●●●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●●●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●● ●●●●●●●●● ● ●● ● ●●●●●●● ●●●● ●●●●●●●●●●●●● ●●●●●●●●● ●●●●●●●●●●●●●● ●●●●●●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ● ●●● ●●●●●●● ●● ●●●●●●●●●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●●●●●● ●●●●●●●●●●● ●●●●●●●●●●●●●●●●●● ●● ●●●●●●● ●●●●●●●●●●●●●●●●●●●●● ●●●●●●● ●●●●●●●●●●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●●●●●●●●●●●●●● ●●●●●●●●●●●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●● ●●● ●●●●●● ●●●●● ●●●●●●●●●●●●● ●● ●●●●●●●●●●● ●●●●●●●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●● ●● ●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●● ● ●●● ●●●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●●●●●●●●●●●●● ●●●●●●●●●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●●●● ●●●●●●●●●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●● ●●●●●●●●●●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●●●●●●●●●●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●●●●●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●●●●●●●● ●●●●●●●●● ●● ●●●●●●●●●●●● ●● ● ●●●●●●● ●●●●●●●●● ●●●●●●●●●● ●●●●●●●●●●●●●●●● ●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ● ●●●●● ●●●●●●●●●●●●●●●●●●● ●●●●●●●●●●● ●● ●●●●●●●●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●● ●●●●●●●●●●●●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●●●●● ●●●●●●●●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●●●●●●●●●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●●●●●●●●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●● ●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ● ●●●●●●●●●●●●●●●●● ●●●●●●●●●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●● ●●● ● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ● ●●●●●●●●●●● ●●●●●●●●●●●●●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●● ●●●●●●●●●●●●●●●●●●● ●●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●●●●●●● ● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●●●●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●● ●●●●●●●●●●●●●●●●●●●●●● ●●●●●●●●●●●●●●−0.4 −0.2 0.0 0.2 0.4 0.6 0.8 1.00.0000.0050.0100.015Log real price (FHFA index) at t relative to 2010,q4Originations at t/Eligible OwnersNotes: Observations are metropolitan area by year-quarter, 2000 through 2010. Horizontal axis: real home price changefrom quarter plotted to 2010, quarter 4, from FHFA repeated sale data. Vertical axis: HECM originations by metropolitanarea and quarter divided by household head homeowners over age 65 in the 2010 census.Figure 3.2 plots a measure of FHA liability for all HECM loans as of the last observation (theearlier of loan termination or quarter 4, 2011). This measure equals the greater of zero or theratio of maximum allowable balance to home value (based on initial appraised value and FHFAmetropolitan price index growth) minus one. This is an estimate of the “moneyness” of the bor-rower’s put option. The circles plot unweighted means by quarter of origination. The line plotsmean moneyness for loans by quarter, with metropolitan area weights governed by the area’s share63of all loans originated through the first quarter of 2000, rather than by time-varying shares.16 Shiftsin metropolitan area origination shares across time account for a strikingly large fraction of FHAinsurance obligations.Figure 3.2: Average put “moneyness” by origination quarter●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●1995 2000 2005 20100. of originationAverage Put Moneyness● Average put moneynessAverage put moneyness, loans weighted by 2000 metro origination shareFigure 3.3 plots by year the number of HECM loans originated in all states (in thousands, circles)and the percentage originated in the “Sand States” of Arizona, California, Florida, and Nevada(line). Sand State markets uniformly witnessed larger price increases and more severe housingbusts than other states, and their HECM originations share jumped during the price and originationboom between 2004 and 2007.16The circles measure∑m∑i∈mtzitNmtNmtNt, whereas the line calculates∑m∑i∈mtzitNmtNm2000N2000, where i ∈mt is a loan originated in metropolitan area m at date t, zit is the moneyness measure, and N is a sum of originatedloans.64Figure 3.3: Sand State origination share by origination quarter● ●●● ●● ●● ● ●●●●●●●●●●●●1995 2000 2005 2010020406080100date● Originations in thousandsSand State Share of OriginationsA regression of originations on past price appreciation and subsequent price declines may be ofinterest, although we do not view such a regression as a way to identify a causal role for put optionvalue in generating adverse selection. Expected put option value and liquidity demand are pre-sumably both unknown functions of multiple lags of price growth. High price appreciation mayhave made HECM tempting for liquidity reasons because they raised the ratio of housing to otherwealth, or because they signalled high put option value due to mean reversion in prices. A regres-sion that shows past appreciation explains away some of the relationship between price declinesand originations does not tell us which of those two channels causes the change in coefficient. Ifthere is no change in coefficient, we still cannot rule out the possibility that some other factor, such65as relatively intense mortgage marketing during the boom generated both the HECM originationsurge and price crashes found in certain markets.We estimated a regression with metropolitan area × year-quarter as an observation, and the ratioof originations to mean metropolitan originations between 2003, quarter 1, and 2008, quarter 4 asa dependent variable.17 We measure lagged capital gains as the ratio of real metro-date-specificFHFA repeated sale price index to the real index value in 2000. We measure the home price bustas the real price ratio between the quarter in question and 2010, quarter 4. This measure is anincreasing proxy for put option value. Conditional only on dummies for year-quarter, we find asignificantly positive coefficient of .35 on the bust measure. When we control for the measureof lagged real appreciation, that variable has a coefficient of .15, and the ex-post bust measure’scoefficient falls to .24. We can reject that the coefficient on the bust is the same with and without theboom control, but the bust coefficient remains statistically different from zero at a high confidencelevel.3.3.2 Estimation FrameworkWe want to know if borrower behavior is consistent with adverse geographic selection havingbeen driven by intent to exploit the put options embedded in HECM loans. Statement 1 abovesuggests that we can learn about intent by asking whether borrowers appear to use all credit justbefore termination when credit available is far in excess of collateral value. If credit were notcommonly exhausted, and draws were commonly evenly distributed across the loan’s life, then wewould expect to see borrowers whose loans terminate with homes worth less than the maximumbalance grabbing all credit just before death. As discussed in Section 3.2, these loan features arenot common, so we take a different econometric approach.17Logs are not feasible as there are many zero origination observations. Note metropolitan area fixed effects wouldpreclude regressions with both the boom and bust measure due to perfect collinearity.66The intensive use of credit near origination and light use thereafter described above could reflecteither (a) preemptive exercise of put options expected to land in the money based on fear of in-capacity coming by surprise before termination, or simply (b) impatience to spend home equity.There is evidence to support impatience as a source of initial credit exhaustion. The most frequentlydescribed uses of credit among HECM borrowers surveyed in Redfoot et al. (2007) are paying offpre-existing mortgages (20%) and home repairs and improvement (18%). Emergency and healthuses combine to 14%. “Improve quality of life,” which could reflect immediate or smooth con-sumption is cited as the primary use by 14% of surveyed borrowers and “everyday expenses” by10%. Thus large expenditure needs at origination are a more common reported use of funds thanslow consumption smoothing. No one in the survey appears to have cited “spend more money thanmy home will be worth before I am unable to do so” or similar use, but such an option may nothave been provided and borrowers might have been reluctant to volunteer such information if itwere.Suppose, then, that there are two groups of metropolitan areas. The “treated” one contains someborrowers of a particular demographic group x who take on loans at date 0 and terminate at date T ,and wind up with credit limits greater than collateral value. In the “control” or “comparison” groupof metropolitan areas, borrowers of type x hold options that are out of the money at T . Assume fornow that liquidity-based credit demand for credit exhaustion through T is identical conditional on xand the different ex-post price realizations; we attempt to verify the plausibility of this assumptionbelow.A fraction µ of the treated borrowers were motivated to take on HECM in part because of ruthlessanticipation of put option value. We assume µ is a population average independent of terminationdate. Suppose that among the ruthless µ, a fraction lµ would exhaust credit prior to T for HECM’s“intended” liquidity purposes even if the put terminated out of the money. Of the µ [1− lµ] ruthlessborrowers who do not exhaust credit for liquidity reasons, suppose that α exhaust credit prior to67termination in anticipation of an in the money termination because they are concerned that if theydo not draw credit early, they will become incapacitated without warning later. Suppose that ofthe remaining ruthless borrowers µ [1− lµ] [1−α] only a fraction k retain the acuity to exhaustcredit. We are unaware of any medical evidence on the correct value of k, the fraction of borrowerswho would be capable of exhausting credit at nearly the last minute if they found it optimal todo so.18 We regard 50% as a plausible upper-bound value for the product [1−α] [1−k], whichrepresents the fraction of ruthless borrowers who become dead or incapacitated without arrivingfirst at a belief that such an event was sufficiently likely as to make the cost of failing to exercise aput option at all greater than the capital cost of failing to wait an additional period.With respect to the remaining 1−µ fraction of borrowers, call l0 the fraction who have no ruthlessintent, but have sufficient demand for HECM for liquidity purposes as to exhaust credit. Somefraction of borrowersψ [1−µ] [1− l0]might exercise put options strategically once stumbled upon.Because an absence of ruthlessness is the more economically interesting result, we bias our implicitestimates of µ upward by attributing any opportunistic credit use to ruthlessness and assumingψ= 0.We regard a plausible assumption to be that l0 > lµ. That is, liquidity demand for HECM is nosmaller among those who take on HECM solely for liquidity purposes than it is among borrowerswho use HECM with the intention of exploiting the put option. This need not be true: a borrowerwith foresight in 2006 might have expected to earn returns much greater than the HECM interestrate by purchasing bonds or taking short positions in stocks. The large number of borrowers whodestroy put option value by using funds to improve their homes, combined with the absence of bor-rowers who state a use of return rate arbitrage, suggests to us that this is a reasonable assumption.We assume the inequality is strict and that liquidity demand for both groups is l.18Causes of death do not seem like a promising source of calibration. Death from heart disease, the leading cause ofdeath, for example, need not arrive as a result of an unanticipated heart attack.68Total credit exhaustion among borrowers with demographic characteristics x treated by an antici-pated in-the-money option at T is then:YxT = l+[1− l]µ [α+[1−α]k] . (3.4)Among the control borrowers of type x and liquidity demand l, a fraction α?µ? might have exer-cised in ruthless anticipation of moneyness at termination. Through date T , this expected money-ness has not been realized. Credit exhaustion in the comparison group through T is:Y?xT = l+[1− l]µ?α?. (3.5)Subtracting (3.5) from (3.4) yields the difference in credit exhaustion between the two groups:YxT −Y?xT︸ ︷︷ ︸difference in credit exhaustion= [1− l] [αµ−α?µ?]︸ ︷︷ ︸Difference in premptive put exercise+ µ [1−α]k︸ ︷︷ ︸Put exercise near termination . (3.6)Equation (3.6) implies that comparing markets with in and out of the money options:Result 1 Suppose that the following conditions hold:1. The ruthless fraction µ is independent of termination date T .2. The fraction of borrowers with liquidity demand for exhausting credit, l, is identical forborrowers of type x in both markets and less than 100%.3. The fraction of borrowers who ruthlessly exercise put options preemptively near originationin expectation of moneyness in the out of the money market µ?α? is strictly less than one andweakly less than µα.Then if the fraction of ruthless in the money borrowers is positive, the difference in credit use amongborrowers with characteristics x terminating at T between these areas YxT −Y?xT is also positive.Assumption 3 should not be controversial assuming adverse selection was driven in part by ruth-lessness. As time progresses between 0 and T , borrowers in the “in-the-money” metropolitan areas69see larger price declines than those in the comparison areas, and so have greater incentive to pre-emptively exercise.Equation (3.6) and Result 1 provide the basis for the regression analysis that follows. If we canidentify appropriate treatment and control groups, a regression of loan exhaustion on an indicatorfor whether or not the loan terminated with an in-the-money put option measures the left-hand-sideof (3.6). If we find tightly estimated differences in credit exhaustion close to zero, the ruthlessfraction of borrowers must also be close to zero. A possibility that we do not wish to rule out isthat we could find a negative value for (3.6) even with a positive fraction of ruthless borrowers µ.This could occur if assumption 2 were violated: some borrowers might not only not ruthless, butmight be actively reluctant to borrow more than they owe, so that liquidity demand is weakened byput option moneyness. We discuss this possibility again in the Conclusion.Under the stronger additional assumptions that α and k are known, and that there is no preemptiveput exercise in the comparison out of the money market: α?µ? = 0, we could solve (3.6) for µ.Notice that if there is no preemptive exercise in the comparison market, then Y?xT = l:µ=YxT −Y?xT[1−Y?xT][α+[1−α]k]. (3.7)Given the strength of the required assumptions, this structural interpretation is only interesting ifwe find a significant “reduced form” difference in credit use (3.6).It may be helpful to compare credit use after the first year of a loan’s life. Given greater demand forHECM loans on the “extensive margin,” we might expect that liquidity demand would also havebeen greater on the intensive margin in the first year of the loan’s life in the markets where putoptions subsequently came into the money. Alternatively, given a lower origination rate, borrowersselecting into HECM in the control metropolitan areas we consider might have had abnormallyhigh initial credit demand. To address these possibilities, in some specifications, we condition on70initial credit draws and ask if put option moneyness is associated with additional credit use in lateryears of loans’ lives. As long as µ is positive, we should find a positive relationship between creditexhaustion and put option moneyness in such a regression. This approach improperly controls forpreemptive put exercise in the first year, but none of the loans we considered featured put optionsin the money until after that year had ended.Another difference of interest is between borrowers in markets treated with in-the-money-optionswho terminate at some date T prior to the end of our sample and those borrowers whose loanssurvive to the end. If µ and l were plausibly identical across those groups, then the difference incredit exhaustion should be µ [1−α]k: the last minute put option exercise among those with loansthat terminate. Whether liquidity demand l up to T would be larger or smaller among terminatedor unterminated loans is not obvious.The theoretical set-up outlined in this section recognizes that the HECM contains an embedded putoption. While our empirical model does not explicitly explore the stochastic nature of house pricesand interest rates on the exercise of this option, we are careful not to claim we solve for optimalcredit use, a spectacularly difficult problem. Instead, we confine the sample to loans that haveterminated. Among terminated loans, if the borrower had the wherewithal to optimally exercise,the solution is trivial: “take the money and run.”3.3.3 HECM Loan MicrodataWe use data on HECMs originated between 1989 and 2011 that FHA has made public (with nowarranty as to accuracy).19 This data includes the date of origination, the mortgaged property’sstate, metropolitan area, Zip Code, and appraised value at origination; the borrower’s age (we recent data has not yet been made public.71the younger age if a couple), marital status (and gender if single); the initial loan amount. Weobserve the date of termination, if any, and an indicator for whether the loan terminated as anFHA repurchase (lenders may force this before termination if the outstanding balance is close toor above initial property appraised value), an FHA insurance claim (available to servicers aftertermination, if collateral value is less than the outstanding balance), or neither. We also observecredit use (or, rarely, a negative draw as partial prepayment) for each year of the loan’s life, throughmid-2011.We are interested in the behavior of lines of credit with credit limits in excess of property value andsimilar lines of credit with no such moneyness. We thus narrow consideration to loans originatedin 2006 and 2007. It can be seen in Figure 3.4 that loans originated thereafter were less likely tohave “crossed over” into moneyness prior to the end of our sample.72Figure 3.4: Distribution of loans terminating “underwater” by quarter of originationllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll1995 2000 2005 20100. of originationCumulative distribution of loans last observe with limit > collateralNotes: We classify a terminated loan to be “underwater” if the estimated principal limit is greater than the estimatedhome value.Also, the introduction in 2008 of fixed rate loans that require 100% credit use at origination isproblematic. There is no interesting variation in credit use in fixed rate loans since prepaymentis so uncommon. Taking a line of credit starting after mid-2008 involved selecting out of a prod-uct that requires credit exhaustion. The Sand State share of originations declined after 2007, sowe are able to capture the peak of adverse selected loans even with the upper bound originationrestriction.Data for the full FHA sample, narrowed to include only loans originated before the end of 2008,73with an observable metropolitan area home price series back at least as far as 1989, with bor-rower age between 62 and 100, with initial credit use below zero and above 100% recoded to thosebounds, and confined to credit line mortgages with gender or marital status reported, are summa-rized in Table 3.1.Table 3.1: Summary statistics of sampleVariable Obs Mean Std. Dev Min MaxOrigination quarter 340,273 2004.74 3.247 1991 2007.75Borrower age 340,273 73.494 7.266 62 124Initial credit to appraised value 340,273 0.632 0.138 0 1Estimated credit to value at last obs. 340,273 0.909 0.363 -0.126 3.04First year credit use to initial year limit 340,273 0.685 15.351 -7876 3545Second year credit use to second year total limit 340,273 0.05 1.111 -636 40Third year credit use to third year total limit 340,273 0.033 1.441 -831 38Fourth year credit use to fourth year total limit 340,273 0.023 1.75 -1017 25Fifth year credit use to fifth year total limit 340,273 0.013 1.806 -997 16Use to avaialble credit at last observation 340,273 0.825 21.968 -11945 3772Maximum ratio: refinance proceeds availableto credit available over loan’s life 340,273 1.149 0.352 1 81.48Appraisal at origination 340,273 232,621 147,936 13,300 950,000Financial Freedom originator or sponsor 340,273 0.375 0.484 0 1We exclude loans originated before 2006 from regression analysis because they seem like a poorcomparison group for borrowers facing in-the-money put options. Loans originated through 2005are unlikely to have crossed over into moneyness, and generally experienced large price increasesafter origination. That appreciation may have generated pressure for increased liquidity demandafter closing and for rapid termination through refinance or sale. We thus view loans with earlierorigination within the sample metropolitan areas to be a less than optimal control group and likelyto overstate pure liquidity demand among treated borrowers Y?. Our approach does not permit es-timation of metropolitan area fixed effects but we suspect loans originated in the Sand States whilehome prices were rapidly rising would make a poor comparison group for the later vintage SandState loans we consider.20 Over 40% of all HECM loans that have terminated with a shortfall claim20Per Davidoff (2014), the share of borrowers living in predominantly black and Hispanic neighborhoods grew more74and were originated before 2008 were originated in 2006 and 2007. We exclude from regressionsa small number of loans with gender or marital status missing, with a younger borrowers’ age over95, or with loan to value ratios at origination that may be outliers (the top and bottom half-percent).Put option moneyness: resale value and available creditWe want to approximate the outstanding balance, credit limit, and mortgaged property value throughthe life of the loan as accurately as possible. We accumulate the outstanding balance and credit limitmonthly at the monthly 1-year treasury rate (taken from the St. Louis Federal Reserve Bank) plusthe loan’s margin rate plus the 0.5% annual mortgage insurance premium. We cap a very smallnumber of outlier loan margin rates at 3.5%. We observe credit use on an annual basis; given heavycredit use at origination, we assume all draws occur at the beginning of the year.Our estimated ratios of outstanding balance to credit available appear sensible. We exclude twoitems in the FHA data from the balance available for cash draws: the initial mortgage insurancepremium of 2% of the maximum claim amount and a formulaic servicing fee set-aside. Theseassumptions generate a large spike at 100% credit use at origination and borrowers drawing 100%at origination under this calculation are very unlikely to make subsequent draws. The distributionof credit use close to 100% is consistent with good approximation. Of the lines of credit originatedbetween 1989 and 2008 with the required data fields populated, we find that fewer than .1% reportnegative first-year credit use, and fewer than 1.5% report first-year use of more than 102% ofinitial credit. 30% of loans feature initial credit use between 98% and 100% of available credit.Of these, roughly 75% report no further credit use. By contrast, only 30% of loans between 95%and 96% of initial credit used in year 1 report no subsequent credit use, and fewer than 20% ofthose with between 90% and 91% report no subsequent use. Our chief interest is in an indicatorsharply in the Sand States than elsewhere through the housing boom, and credit use in heavily minority neighborhoodsis greater than credit use elsewhere within metropolitan areas.75for credit exhaustion, which is not sensitive to the small number of very large outliers indicated inTable 3.1. Overall, there are some outliers in the data but the data quality seems quite good giventhe possibility of mis-reporting. Furthermore, by 2006-2007 the data collection appears to havematured, for example see the above discussion of the distribution of the ratio of credit use to creditavailable.Our estimate of property resale value at a given date is the original appraised value of the homemultiplied by the ratio of the FHFA home price index for the loan’s metropolitan area in the quarterin question divided by the FHFA index at the quarter of origination, less 5% of value in approx-imate selling costs (the amount suggested in Mortgagee Letter 2008-38.) Unfortunately, FHAreports neither the mortgaged home’s resale value nor the insurable difference between collateralvalue and outstanding balance at loan termination. However, we have several ways to check dataquality.Insurance claims are one source for data quality checks. Among loans originated in 2006 and 2007,under 800 of roughly 48,000 that have terminated involve a repurchase, but 4,903 report a shortfallinsurance claim.21 We hope to find that loans that we estimate eligible for insurance coveragebecause collateral is worth less than the outstanding balance do, in fact, report insurance claims,and that other loans do not. False positives (no claim, but we estimate “underwater”) could arisethrough lenders’ choices, FHA rejection of claims, classical measurement error, or worse, classicalmeasurement error combined with selection. A selection problem would arise if loans we estimateto have a principal limit in excess of resale value are more likely to terminate when the principallimit is, in fact, less than true resale value. This would be consistent with incentives: per Davidoffand Welke (2006) once the put option is in the money, remaining in the home does not reducehome equity.21This difference comes from the facts that repurchases (flags 5 and 9 in the data) are triggered by long durations andhigh interest rates, whereas insurance claims without repurchase (flag 10) come from short durations, low interest ratesand steep price declines. The latter have been more common than the former in the sample period.76Figure 3.5 suggests that some measurement error in debt outstanding or resale value is present, butthat lender behavior is also important. The figure plots the fraction of loans that report a claim byrounded (10% bins) estimated outstanding loan balance to resale value at termination. We see agenerally increasing relationship, but a seeming upper bound on claim rates around 60%.Figure 3.5: Proportion of loans assigned to FHA by estimated Loan to Value0.0 0.5 1.0 1.5 2.0 % Loan to Value at TerminationFraction reporting an FHA insurance claimNotes: Fraction of terminated loans not designated as insurance claims that were assigned to HUD at or before termina-tion, by rounded ratio of outstanding balance to initial appraised value or area cap on claim amount.Consideration of lender behavior leads to Figure 3.6. Financial Freedom had a very large marketshare of HECM loans around the price cycle peak.22 Financial Freedom was a subsidiary of Indy-Mac, a large failed institution taken over by FDIC in 2008. We find that Financial Freedom wasfar less likely than other lenders to make claims, and becomes decreasingly likely to successfully22These Financial Freedom HECM loans are different from the proprietary loans Financial Freedom made. Privatelabel reverse mortgages are not in our data.77claim as the gap between outstanding debt at termination and collateral value rises. FDIC officialsconfirm that the successor entity to IndyMac absorbs losses before FHA.Figure 3.6: Comparison of proportion of loans assigned to FHA by originator0.0 0.5 1.0 1.5 2.0 % Loan to Value at TerminationFraction reporting an FHA insurance claimNot Financial Freedom/Indymac Yes Financial Freedom/Indymac Notes: Fraction of terminated loans not designated as insurance claims that were assigned to HUD at or before termina-tion, by rounded ratio of outstanding balance to initial appraised value or area cap on claim amount.Exclusive of loans on which Financial Freedom is the originator or ultimate investor, we find asignificantly tighter relationship between estimated insurable value and insurance claims. Over75% of non-Financial Freedom loans report claims when the outstanding balance exceeds 125% ofour estimate of resale value. Presumably some other lenders, like Financial Freedom, fail to make78claims despite an outstanding balance greater than available credit. The very steep rise near butless than 100% balance to collateral value suggests that we are more likely to incorrectly deem aproperty as out of the money than to falsely deem a property in the money.23FHFA price index growth may be biased upward relative to HECM collateral growth. Davidoff(2014) shows that shortfall claims are more likely than FHFA-based appreciation would suggest,and that most of this is explained by observable differences in Zip Code demographics and esti-mated price movements. We can explore the impact on our data in at least two ways. First, theLocal Initiative Support Coalition (“LISC”) has created Zip-Code level measures of foreclosureactivity relative to metropolitan area averages. These measures are taken with respect to the con-ventional mortgage market, and so HECM loans will not affect this measure. We may presumethat price appreciation has been weaker in the high foreclosure Zip Codes than low foreclosureZip Codes during the price bust. Figure 3.7 plots the insurance claim rate (for non-Financial Free-dom and non-third party loans) by rounded outstanding balance to FHFA-updated collateral ratioseparately for HECM loans in Zip Codes with LISC foreclosure rates greater than and less thanthe metropolitan HECM sample mean. We see that while the rate of FHA insurance paymentsis somewhat higher on average for loans in high-foreclosure Zip Codes, this difference explainsonly a very small part of the distance between observed insurance rates and 100%. The smalldifference in claim rates based on an observable source of differences in collateral performancewithin metropolitan areas suggests that unobservable differences are not likely to be so large as tocause large errors in estimated put option moneyness. We address the possibility with instrumentalvariables described below.23Capone et al. (2010) attribute this sub-100% gradient to strategic undermaintenance.79Figure 3.7: Comparison of proportion of loans assigned to FHA by metropolitan area0.0 0.5 1.0 1.5 % Loan to Value at TerminationFraction reporting an FHA insurance claimForeclosure Greater than Metro MeanForeclosure Less than Metro MeanNotes: Fraction of terminated loans not designated as insurance claims that were assigned to HUD at or before termina-tion, by rounded ratio of outstanding balance to initial appraised value or area cap on claim amount.Zillow produces a Zip Code home price index based on a proprietary modified repeated salemethodology. Table 3.3 shows that log price changes between origination and the end of 2011are similar between the FHFA and Zillow index measures for metropolitan areas where put optionstypically in the money. In comparison areas, price declines appear to be more moderate in neigh-borhoods where HECM loans were concentrated. We find that Zillow Zip Code index values arecorrelated with a coefficient of .80 with our FHFA measure of metropolitan level price appreciation.In a probit for the presence of a shortfall claim on estimated ratios of balance to mark-to-marketvalue, when value is computed based on FHFA, the estimated coefficient is 1.71; using Zillow Zipcode indexes, the coefficient is 1.75. When both are included the FHFA-based value has a coef-80ficient of 1.47 (highly significant) and Zillow .33 (marginally significant).24 Because the Zillowindex is missing for most comparison group borrowers, we generally approximate with the FHFAindex.3.3.4 Samples of loans with and without in-the-money put optionsWe aim to compare borrowers whose loans terminate with their homes worth less than their creditlimit to borrowers likely to have similar liquidity demand, but whose loans do not terminate withput options in the money. This difference corresponds to equation (3.6) and should reflect ruthlessput option exercise. Our favored methodology is to compare borrowers experiencing similar priceappreciation prior to the home price peak, but divergent price movements thereafter. To constructthese two groups, we wish to satisfy three criteria: a large difference in price movements after thecycle peak, a high degree of similarity in price appreciation through the peak, and a reasonablylarge number of observations of each type.We emphasize close matching on home price appreciation prior to origination because a regressioncontrol approach faces identification problems noted in Section 3.3.1. Lagged appreciation maysignal expectations of future price declines, particularly within the treated group. A comparisongroup with very different price booms from the treated group would likely have lower liquiditydemand based on capital gains pressure, but a control for lagged capital gains may bias downwardthe coefficient on ex-post put value from the true effect of anticipated put option value.Our sample construction problem is thornier than the standard propensity matching problem. The24The relative unimportance of within-metropolitan Zip Code differences does not extend past the treated and compar-ison metropolitan areas. A surprisingly large fraction of loans in markets that saw moderate price cycles feature shortfallclaims, and Zip Code characteristics are highly correlated with shortfall claims in those markets. This does not presentus with useful variation due to the very strong correlations among credit demand at origination, Zip Code minority share,and price bust magnitudes. See Davidoff (2014). Based on Zillow data, some heavily minority neighborhoods aroundWashington, DC, and Baltimore appear to have had more mild price declines than metropolitan averages, feeding therefinance patterns below, and explaining the deviation of our sets of metropolitan areas from others in terms of collateralperformance in minority neighborhoods.81“treatment” of put option moneyness is not observed, but rather approximated based on price de-clines, age, and loan-to-value ratio. Also, some borrowers who are truly out-of-the money at theend of our sample may have anticipated a move into the money later. As in standard matchingproblems, we face the problem that considering a larger set of comparison metropolitan areas willmake the treated and comparison groups less similar on the dimension of 2000s home price boom.Unlike the standard case, if we err on the side of too much similarity in ex-ante growth our mis-measurement of differences in put option value likely gets worse.Figure 3.8 shows that HECMs were most prevalent in metropolitan areas and at times with highpast appreciation and large subsequent price declines. These observations occupy the upper rightof the plot. The most suitable comparison metropolitan areas, where liquidity demand is plausiblysimilar, would be located in the upper left corner, in metropolitan areas where there were large pastprice increases through 2006, but relatively modest price declines after 2006. Unfortunately, theupper left corner is largely vacant due to the high degree of mean reversion in the home price panel.As we move away from the upper left, we must either tolerate metropolitan areas with significantlylower ex-ante price appreciation, or metropolitan areas with price declines large enough that asignificant number of borrowers may have held or expected to hold in-the-money put options. Forexample, what is a better control for adversely selected Bakersfield, CA, which saw log real priceincreases and decreases of roughly 80%, New York, where log prices rose roughly 55% then fellroughly 25%, or Orange County, CA, where prices rose roughly 75%, then fell roughly 50%. Weare likelier to incorrectly deem a loan out-of-the-money in Orange County.82Figure 3.8: Distribution of HECM origination share and metro area price growth●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●0.0 0.2 0.4 0.6 0.8 log real price (FHFA index) at t relative to 2010,q4Max log real price (FHFA index) at t relative to 2000,q1Notes: Observations are metropolitan areas. Horizontal axis is price ratio of peak to 2010, quarter 4. Vertical axis isprice ratio, peak to 2000, quarter 1. Bubbles are proportional to total originations 2006-2007 divided by originations1989-2004.To form the treated and comparison groups, we use either all or none of the borrowers in eachmetropolitan area with available home price and HECM loan data. We want as much of the vari-ation in put option moneyness as possible to be driven by differences in price appreciation acrossmarkets. Within metropolitan areas, most variation in moneyness comes from differences in loan83to value at origination due to borrower age. Conditional on borrowers not seeing sufficient capitalgains to refinance, the severity of ex-post price declines provide variation in put option value plau-sibly uncorrelated with liquidity demand. Since credit use generally falls with age in the HECMdata, put value variation based on borrower age is unattractive. Close differences in the ratio ofestimated loan-to-value ratios around 100% are an unattractive source of variation in the indicatorfor put option moneyness given that we are approximating both numerator, the denominator, andthe relevant date at which borrowers expect that their loans will terminate with an in-the-moneyoption.To determine the lists of metropolitan areas, we choose three cutoff numbers: a minimum 2000-to-peak (local peak) real price appreciation for the control group, a maximal peak-to-2011 bustfor the control group, and a minimal bust for the treated group. We select the combination ofcutoff values to maximize the difference between the t-statistic for equal means of group homeprice appreciation between 2000 and the post-2000s peak and the t-statistic for equal means acrossgroups in the ratio of peak to 2011 prices. Recognizing clustering, we take each metropolitan areaas an observation, and do not population weight. We believe this is a natural objective to capturethe need for homogeneous booms, different busts, and adequate sample size, but find similar resultswith several alternative methodologies.25The selected metropolitan areas are listed in Table 3.2. We find that the objective t-test statisticdifference is maximized when we use 11 control metropolitan areas and 27 treatment markets. Notsurprisingly, the high appreciation and high put option value metropolitan areas are all located inSand States. The control metropolitan areas are concentrated between Washington, DC and thegreater New York metropolitan area.25An earlier version of this paper assembled the treated and control groups in a differently ad-hoc fashion. The listswere largely similar to those below, but mean appreciation was more different between the two groups and we controlledfor lagged appreciation in all regressions.84Table 3.2: List of treated and comparison metropolitan areasLog ratio of real post-2000 FHFA peak toComparison Metropolitan Area: 2000q1 FHFA 2011q4 FHFAAtlantic City-Hammonton, NJ 0.66 0.38Baltimore-Towson, MD 0.60 0.33Bethesda-Rockville-Frederick, MD 0.65 0.33Edison-New Brunswick, NJ 0.6 0.33Honolulu, HI 0.61 0.18Kingston, NY 0.61 0.27Nassau-Suffolk, NY 0.60 0.32Ocean City, NJ 0.74 0.35Virginia Beach-Norfolk-Newport News, VA-NC 0.6 0.28Washington-Arlington-Alexandria, DC-VA-MD-WV 0.71 0.38Treatment Metro Area:Bakersfield-Delano, CA 0.81 0.83Cape Coral-Fort Myers, FL 0.8 0.91Carson City, NV 0.64 0.81El Centro, CA 0.63 0.81Las Vegas-Paradise, NV 0.63 1.02Madera-Chowchilla, CA 0.78 0.88Merced, CA 0.85 1.13Modesto, CA 0.83 1.04Naples-Marco Island, FL 0.85 0.84North Port-Bradenton-Sarasota, FL 0.73 0.79Palm Bay-Melbourne-Titusville, FL 0.74 0.80Phoenix-Mesa-Glendale, AZ 0.61 0.79Port St. Lucie, FL 0.79 0.88Punta Gorda, FL 0.72 0.81Reno-Sparks, NV 0.62 0.88Salinas, CA 0.77 0.89Sebastian-Vero Beach, FL 0.73 0.79Stockton, CA 0.78 1.02Vallejo-Fairfield, CA 0.74 0.93Yuba City, CA 0.81 0.89All others (median): 0.19 0.19Table 3.3 provides certain descriptive characteristics separately for the two selected groups ofmetropolitan areas and for all other metropolitan areas. A crucial statistic is that among the loansterminated in 2009 through 2011 that we consider, the median principal limit to estimated resale85value ratio was .78 in the comparison and 1.42 in the treated areas; in other metropolitan areas, themedian was .81.Table 3.3: Characteristics of loans in treated and comparison metropolitan areas:Macro variables: metropolitan areas Comparison Treatment OtherMedian real FHFA Price 2000s peakPrice 2000q1 1.83 2.09 1.21Median real FHFA Price 2000s peakPrice 2010q4 1.39 2.26 1.22Total Ratio of originations 2006-2007 to 1989-2002 2.30 6.40 2.40Total Ratio of Housing Units 2009 to 2000 (US Census) 1.08 1.26 1.12Loan characteristics for 2006-2007 originations Comparison Treatment OtherOriginations with available data 19,731 36,413 107,406Median appraised value at origination 305,590 252,710 200,707Median Age of borrowers 2006-2007 73 72 73Median Age of borrowers terminating 2009-2011 76 77 77Loan-to-value capped by metro-time maximum 0.44 0.34 0.30Mean Zip Code fraction black 0.26 0.06 0.15Mean Zip Code fraction Hispanic 0.06 0.20 0.14Fraction single men 0.17 0.18 0.18Fraction single women 0.50 0.40 0.46Financial Freedom originator or sponsor 0.38 0.35 0.46Median terminal principal limit to value among loans terminated ’09-’11 0.78 1.40 0.81Mean maximal refinance to initial credit over loan’s life 1.03 1.02 1.03Loans originated 2006-2007 terminated prior to 2011q4 0.21 0.11 0.17Loans originated 2006-2007 terminated 2009q1 to 2011q4 0.14 0.07 0.11Apparent refinance??? rate among loans terminated 2009-2011 0.33 0.05 0.18Certain loans? with shortfall claim 0.75 0.12 0.34Loans with available creditestimated value > 1 if held through 2011q4 0.29 0.96 0.39Loans with year 1 credit use > .95 0.37 0.42 0.43Loans with credit use by min(termination, end 2011) > .95 0.57 0.64 0.62. . . among borrowers terminating 2009-2011 0.51 0.54 0.52Mean nominal log FHFA price change origination to end of 2011 -0.28 -0.76 -0.34. . . among borrowers terminating 2009-2011 -0.29 -0.77 -0.32. . . among borrowers with available Zillow index -0.29 -0.76 -0.37Mean nominal Log Zillow price change origination to end of 2011 -0.13 -0.74 -0.26. . . among borrowers terminating 2009-2011 -0.17 -.074 -0.33The differential growth in HECM originations in metropolitan areas with comparable booms butvery different bust magnitudes supports the notion of prescient borrowers put forward by Shan(2011) and Haurin et al. (2014). The ratio of HECM loans originated in the peak years of 2006 to862007 to pre-price spike originations in 1989 through 2003 is 2.3 for the comparison group, 6.4 forthe treated group, and 2.4 for all other metropolitan areas with any HECM originations in 2006-2007. By contrast, across metropolitan areas by group, the mean ratio of real FHFA index to itslevel in 2000, quarter 1 was 2.09 in the treated group, 1.83 in the comparison group, and 1.21 inother metropolitan areas. The median ratio of peak real FHFA price to price in 2010 quarter 4 is2.26 in the treated group, 1.39 in the comparison group, and 1.22 among other metropolitan areas.That the larger price declines in the comparison group might have been anticipated by potentialborrowers is suggested by the ratio of total housing units (based on census counts) in 2009 to 2000across the three areas: 1.27 in the treated Sand State markets, 1.08 in the comparison group, and1.13 among all other metropolitan areas. That is, the treated metropolitan areas had larger supplyelasticities, calling into question the durability of price booms there.Borrowers’ ages across the groups are similar, with means of 73 in the comparison group and 72in the treated group. The treated metropolitan areas have fewer single female borrowers and moresingle men and couples. Appraised property values are higher among the comparison group, anda larger fraction of those borrowers have loan-to-value ratios constrained by time- and location-varying caps on insurable value.26Davidoff (2014) shows that the black and Hispanic population share of homeowners’ Zip Codes(we do not observe borrower race or ethnicity) and the median value of owner occupied homes asof 2000 are strong predictors of HECM participation and credit use conditional on participation.We find that near the peak, the comparison and treatment group borrowers lived in Zip Codes withsimilar sums of black and Hispanic residents, but with the comparison group more black and thetreatment groups more Hispanic.26 Unfortunately, among the treated group, the caps are close enough to appraised value that only a small fractionof treated but capped borrowers terminate with credit less than estimated net resale value. Also, high property valuesare associated with limited credit use even among borrowers not facing caps. It is thus not feasible to compare creditbehavior among borrowers in areas that observed large price declines who did and did not see credit exceed propertyvalue before an actual or likely termination.87Given the difference in the magnitude of price declines, it is not surprising that loan terminationsare less common among the Sand State “treated” borrowers than in the comparison group. In bothgroups, terminations were much more rapid in the years of rapid price appreciation. Of loansoriginated in 2002-2003 in the comparison metropolitan areas, 40 percent had terminated by mid-2007; in the treatment metropolitan areas, 62% of earlier borrowers had terminated by mid-2007.27Table 3.3 reveals that only 13% of loans in the treated metropolitan areas terminated prior to theend of our sample period in mid-2011.One concern we do not directly address when constructing the treatment and control groups isthe possibility that borrower location is endogenously determined. There is some evidence in thecommercial and residential mortgage literature that borrowers propensity to default is differentacross geographic regions. However, this is unlikely to be a factor because the default rate in thecomparison cities (mostly East Coast locations) was lower than in the treatment areas (mostly sandstates). One way we address the issue of propensity to exercise the default option is by confiningthe sample to zip codes with high levels of education.3.3.5 Plausibility of Identifying AssumptionsWe now review the plausibility of the three assumptions that Result 1 lists for the difference incredit exhaustion between treated and comparison borrowers to reflect ruthless credit use.A1. Within the treated metropolitan area, the ruthless fraction µ is independent of termina-tion date T .A significant concern is that borrowers in the treated areas not select into termination becausetheir homes are in fact worth more than their credit limits or because they are less ruthless. Suchbehaviour would bias our estimate of the relationship between moneyness and credit exhaustion27These termination differences are discussed in more detail in Davidoff and Welke (2006).88towards zero if some fraction of the population did use credit ruthlessly.While we see clear evidence of selection across the treated and comparison groups into termina-tions, it is not at all clear that those whose loans terminated within the treated group saw smallerprice declines than those whose loans did not terminate. Absence of such selection may be consis-tent with optimizing behavior if the variance of price outcomes is not too great within groups. Oncea home is worth less than the available balance, there is little reason to leave, so we might expectrandom health shocks and death to generate terminations. Moreover, not all borrowers stood to loseput option value by terminating their loans before the end of the FHA sample period. Since the endof our sample, home prices have risen considerably faster in some of the treated metropolitan areasthan HECM loan balances have grown.28HECM data provide insights into the endogeneity of terminations in different ways. First, wewould like to see that borrowers who we see with large ratios of outstanding balances to estimatedproperty value generate shortfall claims. Otherwise, we would suspect endogenous termination inresponse to options being out of the money. Consistent with evidence in Figure 3.5, among treatedmarket borrowers taking a very large initial draw (greater than 95% of initial value), outside ofborrowers excluded from Figure 3.6, 76% terminated with a shortfall claim. A lender’s shortfallclaim is a sufficient, but not necessary condition for no remaining equity. We thus have reason tobelieve that a large majority of borrowers terminating with estimated credit available greater thanresale value actually did hold in-the-money put options. By contrast, only 16% of this type of loanfeatured a shortfall claim in the comparison metropolitan areas.Refinance behavior also informs the extent of strategic termination. Falling 10-year interest ratesand rising home prices allow borrowers to access more through a refinance than is available on their28A borrower in Phoenix who wished to trade up in home quality, for example, would have been better off purchasingand moving into a new home in 2011 than 2013 (prolonged exit requires a termination). Terminations are likely involun-tary and related to mortality or severe disability among those we deem to have more available credit than property value,since there is little financial incentive to leave (beyond defaulting on property tax obligations).89existing credit line. If a large fraction of loans in the treated areas terminated with a seeming refi-nance, we would worry that there was selection into termination based on idiosyncratic collateralperformance. While a loan record does not indicate the reason for termination, the record does statethe reason for origination. We can thus check for each termination to see whether a new loan hasbeen originated in a matching Zip Code to a borrower with a matching age profile where the newloan is flagged as a refinance. We find that of the loans that terminated after 2009 in the treated ar-eas, approximately 5% terminate with a matching refinance origination. By contrast, roughly 33%of the comparison groups terminations feature such a matching refinance. This provides comfortthat only a small fraction of terminations are voluntary in the treated metropolitan areas. Davidoffand Welke (2013) show that refinances are more procyclical than termination through sale. Weaddress problematic refinances among comparison group borrowers below.We can also ask whether there has been selection into loan termination based on observed FHFAmetropolitan or Zillow Zip Code price index appreciation since origination. Table 3.3 shows that inthe treated metropolitan areas, mean FHFA log appreciation between origination and 2011 quarter4 is very close to equivalent among those terminating their loans after 2009 and others. Among alltreated borrowers, mean log appreciation between origination and the end of 2009 is -.77 for FHFAmetropolitan growth and -.74 for Zillow growth. For those who terminate after 2009, the meansare -.76 and -.74. The high degree of similarity in appreciation rates between terminated and un-terminated loans, the low rates of refinance among all terminations, and the high rates of insuranceclaims on high-initial draw loans that terminated all suggest that loans in the treated metropolitanareas that terminated after 2009 were, indeed, likely to feature in-the-money put options. In thecontrol group, mean log FHFA metropolitan appreciation between origination and the end of 2011is -28% among all loans and -29% among those that terminated between 2009 and 2011. Amongcontrol group borrowers, the difference in log Zillow Zip Code index between origination and De-cember, 2011 among loans terminating before 2009 or not at all is -.13%, as opposed to -.17%90among all loans.Finally, we regress an indicator for having terminated after 2009 on various borrower and loancharacteristics for both sets of metropolitan areas. The results are summarized in Table 3.4. Wefind that conditional on the date of origination, the change in the FHFA index between loan origi-nation and the end of 2011 is negatively associated with terminations in the treated group. We findno relationship between the probability of an exit and Zillow price appreciation. Zip Code fractionnon-Hispanic white and initial credit use also fail to predict exit. These facts suggest that ruth-less borrowers do not seek to avoid exit as put option value grows, leaving a non-representativelyruthless sample of terminated treated loans.Specifications (3) through (6) of Table 3.4 confine the analysis to the comparison metropolitan ar-eas. In (3) and (4), we find results broadly consistent with those among treated borrowers. FHFAprice growth, initial credit, and Zip Code non-Hispanic white share are not associated with exit.However, exit is correlated positively with Zip Code price growth. Specifications (5) and (6) ad-dress selection into refinance versus non-refinance termination. While the low rate of refinance intreated metropolitan areas is a good signal for our identification strategy, the high rate of refinancein the comparison metropolitan areas is worrisome in two ways. First, it is ambiguous whether ornot to treat refinance as a termination; that is, we wish to compare terminated loans in the treatedmetropolitan areas to comparable terminated loans in the control group. Refinanced loans aretreated as a termination in the data, but are not exactly exits from the HECM program. Borrowerswho refinance increase their credit limits but pay additional insurance fees. Second, we might beconcerned that terminated loans are concentrated among borrowers with high liquidity demand inthe comparison metropolitan areas due to appreciating prices.In the comparison group, we find that there are significant differences among borrowers who termi-nated through refinance and those who terminated through death or sale. Unlike in the full sample91of terminated comparison group loans, when refinances are excluded, we find significant positiveselection into non-refinance termination on ex-post metropolitan and Zip appreciation and nega-tive selection on initial credit use. We find that borrowers in Zip Codes with more non-Hispanicwhites are more likely to exit through non-refinance means. Given that the full sample of compar-ison group terminations is less selected on important characteristics, we believe it is appropriateto use this full group as a comparison for terminated treated group loans. It appears that compari-son group borrowers with high initial credit use are no more or less likely to terminate than otherborrowers, but conditional on termination they are relatively likelier to refinance. However, giventhe ambiguity, we also present results using all loans, terminated or not, and only non-refinancedloans.92Table 3.4: OLS regressions of an indicator for loan termination(1) (2) (3) (4) (5) (6)Log FHFA growth origination-2011 -0.060* -0.057 -0.037 -0.163 0.220** 0.124(0.024) (0.032) (0.115) (0.265) (0.042) (0.069)Zip Code % white 0.012 0.013 0.004 0.012 0.044** 0.051**(0.013) (0.013) (0.012) (0.015) (0.009) (0.009)Credit limit to value origination -0.256** -0.210** -0.788** -0.684** -0.363** -0.278**(0.025) (0.040) (0.076) (0.128) (0.039) (0.075)Financial Freedom 0.021* 0.025* 0.031** 0.034** 0.024* 0.026*(0.008) (0.009) (0.007) (0.010) (0.008) (0.009)Third Party -0.011** -0.013** -0.015* -0.014 -0.012* -0.012(0.003) (0.004) (0.006) (0.009) (0.006) (0.007)Initial Credit limit/appraisal -0.009 0.002 0.001 0.007 -0.048** -0.048**(0.008) (0.011) (0.010) (0.010) (0.007) (0.010)Log FHFA growth 2000-origination -0.016 -0.025 -0.026 0.079 0.062* 0.093(0.032) (0.029) (0.036) (0.097) (0.025) (0.052)Log initial appraised value 0.018 0.027* -0.056 -0.047 -0.017** 0.001(0.010) (0.011) (0.034) (0.052) (0.005) (0.013)Credit limit capped -0.026** -0.024** 0.001 0.012 -0.007 -0.004(0.004) (0.006) (0.012) (0.017) (0.005) (0.007)Log Zillow growth origination-2011 -0.003 0.181* 0.073*(0.014) (0.063) (0.027)Constant -0.009 -0.143 1.299* 0.897 0.484** 0.124(0.141) (0.136) (0.443) (0.662) (0.080) (0.173)Adjusted R2 0.05 0.05 0.07 0.06 0.06 0.06Number of Observations 36,399 18,659 19,694 11,090 18,332 10,179Group Treated Treated Comparison Comparison Comparison ComparisonExclude refinance No No No No Yes YesNote: * p < 0.05; ** p < 0.01Further inspection of apparent refinances reveals that they are highly concentrated among borrow-ers in minority neighborhoods who exhaust or nearly exhaust credit in the first year of the loan’slife, and they are commonly originated through third parties into a particular investor, GenerationMortgage Company. The refinanced loans warrant future study.A2. Identical liquidity demand across groups, conditional on characteristicsWe do not observe liquidity demand. One particular concern, though, can be addressed. Both thetreated and comparison metropolitan areas saw price declines after 2006, but they were more severein the treated areas. Thus there may have been greater demand for consuming home equity after93origination in the comparison than treated areas. We ask whether borrowers in metropolitan areaswithin the comparison group responded to price changes in their credit use by regressing credituse through the sample end on borrower and loan characteristics and the rate of price appreciationthrough the sample period end. Table 3.5 presents repeated logit estimates of credit exhaustion foreach year of a loan’s life. We find that within the control areas, with generally out-of-the-moneyput options, relative growth in the FHFA index is associated with insignificantly less propensityto exhaust credit in a given year of the loan’s life. The low rate of insurance claims in this groupsuggests that this is not evidence in the treated group of ruthless exhaustion of credit in the faceof in-the-money put options (a similar result holds with the price growth measure is truncated atthe 50th or 75th percentile). There is thus no evidence that rising prices are associated with greatercredit use. This makes sense in that most borrowers did not see rising home equity over the sampleperiod.A3. The fraction of borrowers who ruthlessly exercise put options preemptively near origi-nation is weakly greater in the treated metropolitan areasThat µα, the fraction of treated borrowers with ruthless intent exhausting credit early in anticipa-tion of put options arriving in the money, exceeds µ?α?, the comparable fraction in the comparisongroup, is highly plausible. HECM originations grew more than twice as fast in treated marketsdespite comparable price growth between 2000 and 2006. If this difference is explained by ruth-less borrowers correctly perceiving greater put value and hence participating more in the treatedmarkets, the condition must be satisfied.94Table 3.5: Panel logit regressions of credit “exhaustion”Metro areas Treated Comparison ComparisonExit 2009-2011 -0.062 0.088 0.089(0.061) (0.057) (0.057)Log price index change origination to date -0.783(0.564)Year 2 -1.277** -1.189** -1.241**(0.041) (0.095) (0.102)Year 3 -1.343** -1.274** -1.391**(0.039) (0.085) (0.117)Year 4 -1.429** -1.332** -1.491**(0.044) (0.124) (0.139)Year 5 -1.811** -1.713** -1.914**(0.054) (0.134) (0.183)Year 6 -2.928** -2.720** -2.493**(0.093) (0.160) (0.258)Log appraised value -1.000** -1.069** -1.073**(0.060) (0.134) (0.135)Limit to value at origination -3.585** -3.272** -3.280**(0.121) (0.139) (0.142)Zip Code % white -0.909** -0.919** -0.913**(0.088) (0.135) (0.133)Constant 15.990* 11.061 10.730(6.339) (10.005) (10.006)N 113,290 63,518 61,968Note: Standard errors clustered at the loan level. All specifications have an age polynomial and gender/marital statusand loan origination month controls. * p < 0.05; ** p < Credit Exhaustion Among Selected BorrowersTerminated loans: treatment vs comparison groupsWe now ask whether terminating a loan with an in-the-money put option is associated with creditline exhaustion. We wish to interpret the difference in credit use associated with moneyness asthe effect of moneyness holding liquidity demand constant. The top panel of Figure 3.9 plotsthe cumulative distribution of ratios of first year credit draws to available credit for loans in thecomparison (line) and treated (circles) metropolitan areas. Both groups exhibit a large mass above95%: 37% among the comparison group and 42% among those treated with a large ex-post price95decline. Among loans terminating after 2009, late enough for treated put options to arrive inthe money, 51% of comparison group borrowers exhausted credit and 54% of treatment groupborrowers. We thus find a small but significant unconditional difference in credit exhaustion thatdoes not rise towards loan termination. If we assume that the fraction of ruthless borrowers ableto exercise the option early or near termination α+ [1−α]k is .5 among treated borrowers, butzero among comparison group borrowers, then equation (3.7) implies a ruthlessness fraction of.54−.51.5[1−.51] = 12.2%. If 80% of those with in the money put options are able to exercise the estimatedfraction is 7.7%.29Recognizing that put option moneyness is correlated with borrower age, and that there are dif-ferences between borrowers in the two sets of metropolitan areas, Table 3.6 presents regressionestimates. The dependent variable indicates a terminal ratio of outstanding balance to credit limitof 95% or greater. We present very similar estimates of the coefficient of interest when exhaustionis defined at 90% use of available credit. The right hand side variable of interest is an indicator forwhether or not the loan terminated with a principal limit in excess of estimated value (appraisal atorigination net of selling costs, inflated by the FHFA metropolitan index). The coefficient on thatvariable corresponds to the value YxT −Y?xT . We are interested in discrete indicators for moneynessand credit exhaustion, because theory does not provide a result of continuously increasing credituse in the ratio of credit limit to property value.29These and other presented point estimates ignore uncertainty in Y? and the fraction of the population capable ofexercising the option. The standard error of the estimate of exhaustion is small enough that convexity of the ratio (3.7)in its denominator has approximately no impact. Uncertainty over k and α would likely matter if we had a distributionof those values in mind.96Table 3.6: OLS regressions of an indicator for near exhaustion of credit (95%+ use)(1) (2) (3) (4) (5)Credit limit > Collateral at Termination 0.021 0.026 0.021 0.005 0.017(0.044) (0.050) (0.071) (0.073) (0.027)Loan to value at origination -0.415* -0.555 -0.414 -0.390 -0.047(0.192) (0.396) (0.220) (0.213) (0.117)Financial Freedom 0.089** 0.140** 0.089** 0.085** 0.043*(0.018) (0.030) (0.018) (0.018) (0.019)third_party 0.013 0.006 0.013 0.014 -0.050**(0.016) (0.020) (0.017) (0.017) (0.018)Log appraisal at origination -0.157** -0.190** -0.157** -0.160** 0.001(0.051) (0.048) (0.051) (0.048) (0.019)constant 2.905** 3.306** 2.904** 2.936** 0.804*(0.740) (0.782) (0.738) (0.709) (0.346)Age and Gender Dummies Yes Yes Yes Yes YesInitial balance controls No No No No YesQuarter of origination dummies Yes Yes Yes Yes YesAllow capped loans Yes No Yes Yes YesInstrument None None Treated metro Claim rate Claim rateAdjusted R2 0.06 0.09 0.06 0.06 0.23Number of Observations 5,214 2,693 5,214 5,160 2,968Mean fitted exhaustion, treated groupif credit were < collateral .54 .52 .54 .52 .47Implicit µˆ per (3.7)µˆ if α+[1−α]k= .5 .09 .11 .09 .02 NAµˆ if α+[1−α]k= .8 .06 .07 .06 .01 NACredit limit > Collateral coefficientUnder 85, high education, income 0.055 0.063 0.049 0.023 -0.036(0.041) (0.052) (0.074) (0.092) (0.062)All loans -0.010 0.027 0.036 0.025 0.037**(0.036) (0.027) (0.049) (0.050) (0.010)Terminated, likely refinance excluded 0.112* 0.122* 0.163* 0.160* 0.042(0.044) (0.049) (0.075) (0.075) (0.032)If cutoff .95→ .9: 0.008 0.007 0.009 -0.003 -0.010(0.043) (0.047) (0.069) (0.070) (0.036)Notes: Credit limit and collateral value are calculated as described in the text. Initial balance controls include indi-cators for ranges of 5% first year credit ranges plus a linear control for first year balance. Capped loans are those withappraised values greater than time- and market-specific FHA caps on insured value. Robust standard errors clusteredat the metropolitan area (CBSA) level in parentheses. Third party loans are those in which a “sponsor” different fromthe “originator” is identified. A large fraction of these loans involved Financial Freedom/ IndyMac. The treated metroinstrument is an indicator for being in one of the treated metropolitan areas. The claim rate instrument is the percentageof loans in the borrower’s metropolitan area with first year balances greater than 95% of initial credit that are flagged ashaving featured a shortfall claim, excluding loans originated by Financial Freedom or third parties. The credit limit >collateral coefficient at a cutoff for exhaustion at .9 rather than .95 is conditional on the same variables as in the mainspecification.97Given a large number of dummy variables, including the regressor of interest, and an instrumentalvariables approach to measurement error, we report results from linear probability model ratherthan probit or logit specifications; such specifications produce similar results. We also find similarresults in repeated probit or logit regressions of the form presented in Table 3.5. We find generallyfewer than five percent, and often less than one percent, of predicted exhaustion rates are outsidethe range 0 to 1. We cluster robust standard errors at the metropolitan level, finding minimal effectson standard errors of clustering also or separately on the date of origination.The main estimates presented in specifications (1) through (5) of Table 3.6 confine the analysis tothe treated and comparison metropolitan areas and consider only borrowers whose loans terminatedbetween 2009 (when loans mostly start coming into the money) and the end of the sample periodin mid-2011. We include control dummy variables for age and gender/marital status, for whetherFinancial Freedom or a third party originated the loan, for the year and quarter of origination,and continuous controls for initial appraised value, and initial credit to appraisal ratio. We do notcontrol for the date of loan termination in the main specification; the treated borrowers terminateroughly six months later on average, so the coefficient on moneyness is biased slightly upwardfrom zero. We include termination (and hence duration) controls in full sample regressions.Estimating the difference (3.6) with controls present assumes that these controls do not affect ruth-lessness conditional on holding a put option in the money. In the case of some variables, that maybe a controversial assumption; for that reason we exclude both price appreciation through the homeprice cycle peak and the black and Hispanic fraction of older homeowners in a Zip Code.30Specification (2) excludes the significant minority of borrowers who faced market-specific loanlimits imposed by FHA, such that loan-to-value ratios are low. We might expect these borrowers30Including these controls decreases the estimated coefficient of interest by roughly .004 in an untabulated specifi-cation, with the controls working in opposite directions. Lagged price appreciation clearly might signal expected putoption value. Given the geography of subprime loan originations, it is possible that borrowers might have anticipatedlarger price busts in minority neighborhoods; see Davidoff (2014).98to behave less ruthlessly than others; their existence in significant numbers should arguably informour estimate of the role of put value calculation as a driver of the Sand State surge in HECMoriginations, though, so we exclude them from only one specification. We find almost preciselyzero difference in the estimated relationship between holding an in the money put option and creditexhaustion whether these loans are excluded or not.Specifications (3) through (5) of Table 3.6 acknowledge that we have an imprecise measure of mon-eyness that is in part endogenously determined by the date of termination. We thus instrument forthe indicator for collateral worth less than the loan limit with two different variables. In specifica-tion (3), we instrument with membership in the treated set of metropolitan areas. In specifications(4) and (5), we instrument with a continuous variable: the metropolitan area-specific fraction ofterminated loans with initial credit use greater than 95% that are not associated with FinancialFreedom and are noted as featuring a shortfall insurance claim. These specifications recognize thatsome borrowers in the comparison group may have held in-the-money options and allow for thepossibility that some in the treated group did not. Neither of these instruments should be caused byendogenous borrower choices after observing put value.31 The rate of insurance claims for loanswith large initial balances will pick up differences in the extent of moneyness. The first stage re-gressions are extremely strong, with F-statistics greater than 300 and t-statistics on the instrumentalvariable above 30, so there is no problem of weak instruments. We do not interact our instrumentalvariables with loan-to-value ratio, because loan-to-value ratio is highly correlated with age, and wedo not believe it makes sense to assume stronger effects for older, deeper in-the-money borrowers.Indeed, we check for such a difference and find no significant difference.In each of specifications (1) through (5) we find a small difference in the prevalence of credit ex-haustion prior to loan termination that is indistinguishable from zero. The IV coefficients are not31The treatment group variable may be correlated with measurement error, since the loan-to-value ratio in excess ofunity indicator includes FHFA price growth, upon which the two groups are selected.99significantly different from each other, from zero, or from the OLS estimates, with a maximumvalue of .026 in specification (4). In that specification, we find a fitted Y? of 52% credit exhaustionbased on characteristics other than put value. Taking the largest IV point estimate of .026, assumingability to exercise [α+[1−α]k] is equal to .5, and using Y? = .52,32 we obtain an implied measureof ruthlessness µ of .057.5[1−.52] = 11%. If capacity to ruthlessly put is .8 our maximal estimate ofµ falls to .07. These values assume that none of the credit exhaustion in the treated metropoli-tan areas reflected ruthless but ex-post generally incorrect preemptive put exercise in comparisonmetropolitan areas.In untabulated regressions, we find an insignificant negative coefficient of -1.2% on terminatingwith a loan-to-value ratio greater than 100% when the FHFA-based price growth estimates throughtermination are instrumented with Zillow Zip Code price growth estimates. There is also no im-pact if we replace the loan-to-value measure at loan termination with the measure if the loan hadsurvived to the last quarter of 2011 (eliminating a correlation between the regressor of interest andloan duration). In sum, recognizing that put value at termination is both measured with error andcorrelated with termination date, does not lead us to a conclusion that ruthlessness drove a large partof the relative doubling in origination growth between the treated and comparison markets.Specification (5) conditions on initial credit use by excluding loans with initial credit use greaterthan 90% of initial balance, and then controlling for indicators for 5% bins of first year credit usealong with a linear term in first year use. We again estimate approximately zero tendency for ex-hausting credit after origination when the principal limit exceeds property value using the insuranceclaim rate instrument. This is a particularly significant finding in light of the possibility that largeinitial credit use in the comparison group may have reflected intent to exploit an underpriced put32We estimate Y? as the mean fitted value of the dependent variable among the treated borrowers, minus the meanindicator for moneyness among those borrowers times the moneyness coefficient. That is, we estimate the rate of creditexhaustion for borrowers with similar characteristics to those with in the money put options, if they did not hold suchoptions.100option that ex-post did not turn out to be underpriced (α?µ?). This specification has the appealof purging differences in liquidity demand between the two groups that might be deemed constantthrough the life of the loan. Given statement 2 in Section 3.2.2, it is not unreasonable to considerterminal credit use conditional on initial credit use to measure ruthlessness. It is not clear theoreti-cally that initially high credit use should be a signal of intent to use HECM as a put option. Indeed,relatively weak demand for HECM’s liquidity features might be taken as a signal interest in putoption value conditional on participation.The bottom panel of Figure 3.9 is a graphical version of specification (5). It is difficult to findevidence that borrowers who failed to use all available credit at loan origination were likelier to usecredit prior to termination in the treated than comparison markets. Of particular interest is the verylow level of credit exhaustion at low levels of initial credit use. For those borrowing less than 50%in the first year of the loan’s life, while a very large majority of those in the treated group (88%,identical to the fraction of all treated borrowers) held options estimated to be in the money, just 12%exhausted credit prior to termination. Assuming a very high strategic failure rate [1−α] [1−k]value of 50%, even if the counterfactual Y? liquidity demand for exhaustion were 0, we wouldestimate µˆ at only .12.5 = 24%. 24% is small relative to the disproportionate origination growthin these markets. In fact, recognizing that credit exhaustion among borrowers in the comparisonmetropolitan areas, a value of µ around zero appears correct among low initial credit use borrowers.Among borrowers with loans terminating after 2009 in the comparison metropolitan areas usingless than 50% of available credit in the first year of the loan’s life, we estimate fewer than 8%terminated in the money. Of this group, over 22% used all credit prior to termination.3333Subtracting 22%× (1−8%) for the comparison group from the 12% low initial credit use among treated borrowersimplies a negative numerator in (3.7).101Figure 3.9: Distribution of credit use across Treated and Comparison metropolitan areasllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll0 20 40 60 80 1000. of Initial Credit Used in Year 1Fraction of Borrowers Using less than this in year onel TreatedComparisonllll ll llllllll lllllll l0 20 40 60 80 1000. Initial CreditFraction Terminating With Oustanding Credit/ Estimated Value > .95 l TreatedComparisonNotes: Treated (circles) and Comparison (lines) metropolitan areas for loans originated 2006-2007. Data points aremeans by rounded hundredths. Left panel: first year credit use divided by available credit. Right panel: fractionof borrowers whose loans subsequently terminated at each rounded five percent initial credit use interval that are lastobserved with credit use in excess of 95% of available credit at terminationThe relative growth between treated and comparison metropolitan areas in loans with initial credituse less than 50% was similar to relative growth among loans with larger initial draws. Considerthe ratio of loans originated prior to 2003 to originated in 2006 and 2007. This ratio is 6.4 to onefor all loans in the treated area, and 4.6 to one for loans with less than 50% initial credit use. In thecomparison areas, the ratio is 2.26 for all loans and 1.6 for loans with small initial credit use.Summarizing, among terminated loans, we can not reject zero effect of high terminal loan-to-valueratios on the propensity to exhaust available credit prior to termination. Point estimates remainclose to zero when we instrument with the treated or untreated status of borrowers’ metropolitanareas, with a metropolitan measure of shortfall insurance claims among loans with intense credituse before price changes were observed, or with Zillow estimated Zip Code price appreciationbetween origination and the end of 2011. The infrequency of credit exhaustion among borrowers102with low initial credit use facing in-the-money put options, the similarity of this exhaustion rate toamong out-of-the-money borrowers, and the similar relative growth of low initial credit loans inthe treated Sand State markets casts severe doubt on intent to put as a cause for selection.Terminated vs. Unterminated Loans Within Market GroupsAn alternative proxy liquidity demand Y? for our terminated in-the-money loans comes from un-terminated in-the-money loans. If we believe that terminations are random conditional on ageand gender, presumably reflecting death (which is difficult to forecast more than one or two yearsahead, see Chalmers and Reuter (2012)), then we may safely assume that both liquidity demandand preemptive put exercise are similar in the terminated and unterminated groups. Table 3.4 sug-gests this assumption may not be far off. We estimate a panel repeated logit (not a hazard, givensmall numbers of draw observations) of the form:ymit = f(α+β1exiti+xiγ,it). (3.8)In equation (3.8), ymit indicates if borrower i in metropolitan area m in the tth year of her loanexhausts credit. We restrict the sample to loan-year pairs in which the borrower did not previouslyterminate their loan or exhaust credit.34 The controls x are: dummy variables for the year andmonth of origination, borrower gender, a cubic in younger borrowers’ age, the fraction of home-owners in the borrower’s zip code as of 2010 who are white and non-Hispanic, and log appraisedvalue at origination. We include dummy variables for the year t of the loan’s life (1-5 in our sam-ple) in which the indicator for credit exhaustion is measured. β1 is the object of our interest: weask if exit from the home after the year 2009 is associated with additional credit use in any yearprior to exit. Given the possibility that borrowers preemptively respond to forecast future termina-34We do not check for a potentially non-zero but tiny number of loans that re-enter non-exhaustion through repaymentof some borrowing.103tion and the fact that exit years are shorter than average, we ask if there is ever a response. Resultsare not meaningfully different if the exit indicator is only turned on in the year of exit.We estimate equation (3.8) among borrowers in each of the two market types in Table 3.5. Specifi-cation (1) confines the sample to borrowers in the treated metropolitan areas. We find an insignif-icant coefficient of -.062 on the indicator for exiting in 2009 or after. There is thus no evidencethat borrowers seek to exhaust credit before exiting their homes, even when their put options arein-the-money. Specification (2) shows of Table 3.5 repeats the analysis in specification (1), but isconfined to the comparison metropolitan areas. We find an insignificantly positive coefficient onexit from the home in these areas with credit exhaustion in a given year as the dependent variable.Standard errors are clustered on metropolitan area only.Treated vs. comparison areas, all borrowersThe facts that terminating borrowers do not use more credit in the treated metropolitan areas butappear to in the comparison metropolitan areas suggest specifications that compare credit exhaus-tion among all borrowers. Per statement 2 in Section 3.2.2, ruthless credit use might or might notinvolve credit exhaustion responding to put option moneyness well before termination. Table 3.6thus includes a row in which each main specification is repeated, but the sample is no longer con-fined to terminated loans only, and we include dummy variables for the date of loan termination orfor having survived through the end of the sample period. Not surprisingly in light of the resultsin Table 3.5, the coefficient of interest, on put option moneyness, is consistently highly similarwhether the sample is or is not restricted to loans that terminated in 2009 or later. We estimate adifference in exhaustion of 3.7% conditional on not having exhausted in the first year and on first-year credit use. This coefficient is in line with others, but statistically distinguishable from zero.104Treated vs. comparison areas, terminated loans only, no suspected refinancesTable 3.6 includes a row that re-estimates specifications (1) through (5) with the sample confinedto loans terminated after 2009 that are not deemed likely refinances. Excluding refinances has alarge effect on the estimated coefficient. We estimate a difference in exhaustion between loanswith and without in-the-money put options as large as 16%. However, conditioning on initialcredit use, which is much larger among apparently refinancing borrowers in the comparison areasthan other terminating borrowers, the coefficient falls to 4.2%, not significantly different fromzero and consistent with other point estimates. Among non-refinancing borrowers, then, treatedborrowers differ from control borrowers primarily in first year credit use. A natural explanationfor the pattern is that high credit use borrowers are inclined to terminate through refinance; byconfining the sample to borrowers who do not refinance, we obtain a sample of control groupborrowers that under-represents heavy first year credit users. This only occurs in the control groupbecause treated group borrowers were almost uniformly unable to execute cash-out refinances dueto falling prices.35 Once we condition on initial credit use, and eliminate the sample selectionproblem, there is very little difference in credit exhaustion between the groups. Conditioning oninitial credit group eliminates this sample selection problem.Older BorrowersRecognizing that there may be lingering concerns about endogenous termination, we also considerboth the possibility that borrowers whose loans terminated were less ruthless than other in-the-money borrowers and that surprise death or incapacity (low αk) might have reduced the number ofborrowers with intent to exploit put options who actually did so. We thus consider borrowers whowere demographically very likely to terminate their loans. In particular, we identify the age andgender combinations that terminated loans at a rate of more than 10% per year among borrowers35The size of refinanced loans is suspicious and warrants future research.105with loans originated in 2006 and 2007, but not living in either of our sample sets of metropolitanareas. We select 10% because interest rates were uniformly less than 5% above the riskless rateon loans. Thus, assuming a loan balance outstanding equal to or greater than appraised value,a 10% probability of death arriving fifty percent of the time without warning would then makeearly withdrawal of all remaining credit profitable even for patient borrowers. If surprise death isa common phenomenon, most of these borrowers should have found early withdrawal profitable.These ages start at 84 for single men, 88 for single women, and 89 for couples. These groupsrepresent a trivial 2% of all borrowers in the treated metropolitan areas, but a larger 9.5% of treatedborrowers who terminated their loans prior to the end of the sample period.A strength and weakness of considering credit use among extremely old borrowers is that these bor-rowers obtain large initial loan-to-value ratios. Thus we can be quite confident that older borrowersin the treated Sand State markets would have held in the money put options as long as terminationoccurred after 2008. However, while only 16% of all comparison group borrowers are estimatedto have been in-the-money if unterminated through 2011, a significant majority of such borrowersin the selected age categories would have been in-the-money (although only a minority of thoseestimated to have borrowed more than available credit are observed with an insurance claim). Wethus do not wish to compare credit use exhaustion among the elderly across markets. Rather, weask if the borrowers demographically likeliest to terminate exhausted credit more frequently thanborrowers who actually terminated their loans in the treated metropolitan areas.Figure 3.10 plots the cumulative distribution of initial credit use in the top panel and exhaustionof credit conditional on initial use in the bottom panel. We separately plot use for all borrowersin the treated metropolitan areas whose loans originated between 2006 and 2007 (circles), thosewhose loans have terminated (solid line), and those most likely to terminate due to demographics(dashed). Older borrowers use less credit at origination, but are similarly likely to exhaust creditthereafter.106Figure 3.10: Distribution of credit use among borrowers in the “treated” metropolitan areasllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll0 20 40 60 80 1000. of Initial Credit Used in Year 1Fraction of Borrowers Using less than this in year onel TreatedTreated, TerminatedTreated, Oldll llll llllllllllllllll l0 20 40 60 80 1000. Initial CreditFraction Terminating With Oustanding Credit/ Estimated Value > .95 l TreatedTreated, TerminatedTreated, OldNotes: All loans (circles), terminated loans (line) and borrowers demographically likely to terminate at more than 10%rate per year (dashed line). Data points are means by rounded hundredths. Left panel: cdf of initial credit use. Rightpanel: fraction using 95%+ of available credit at last observation of the loan (termination or end of the panel in late2011) by rounded five percent bin of first-year credit use.Demographics and Sensitivity of Credit Use to Put Option ValueHECM loans have proven disproportionately popular in neighborhoods with low incomes, and lowproperty values. While we cannot observe borrowers’ financial sophistication, we might believethat sophistication is greater among borrowers with higher property values, living in higher in-come communities and who are relatively young. We consider this possibility by estimating theregressions in Table 3.6, but confining the sample to terminated loans among borrowers under 85in Zip Codes with median incomes and college education attainment rates above the medians inour treated and comparison loan sample. The results are consistent with other specifications; therelationship between put option moneyness and credit line exhaustion are never significantly dif-ferent from zero or from in the main specification; point estimates for the relationship are slightlyhigher, except for use of credit after origination. This pattern does not provide strong evidence thatruthlessness varies with demographics.1073.4 ConclusionFHA has faced costly geographic adverse selection in its HECM insurance program. Amongmetropolitan areas witnessing extremely high price growth between 2000 and 2006, demand growthwas more than twice as great in elastically supplied Sand State markets that saw extreme subse-quent price declines than in the less elastically supplied, mostly Northeastern housing marketswhere price increases were relatively sustained.While many borrowers exhausted their credit lines near the cycle peak, other borrowers do notappear to use credit in the ruthless way that we hypothesize they would if they used HECM toexploit underpriced embedded home price insurance. Close to zero percent of borrowers usedHECM as a pure put option, drawing no credit until after price uncertainty played out, and thenexhausting available credit prior to termination. Among borrowers with credit limits greater thancollateral value, those who terminate after 2009 are no likelier to exhaust credit than those whoseloans survive through 2011. There is thus no evidence that borrowers seek to exercise in-the-moneyput options prior to their expiry.Among borrowers whose loans terminated after 2009, there is a small but significant difference incredit exhaustion between borrowers whose loans terminate with and without credit limits greaterthan estimated property value. However, this difference arises only in the first year of the loan’slife, and becomes indistinguishable from zero conditional on demographics and loan characteris-tics. We find similar differences in credit exhaustion between loans that are last observed with putoptions in or out of the money whether we include or exclude loans still active through 2011 in theregression sample. We find this difference to be significantly different from zero in only one of fivespecifications in the full sample.It is difficult to distinguish ruthless behavior from simple demand for credit when initial credit useis large, but assuming borrowers in high price growth and modest price crash markets used credit108only due to liquidity demand, we describe a way to place an upper-bound estimate on the degreeof ruthlessness among borrowers selecting into the large price crash markets. We find that ourlargest point estimates imply ruthlessness among roughly 12% of treated group borrowers. Thisis not enough to explain the more than doubling of origination growth between the two sets ofmarkets. Conceivably, a large and similar fraction of borrowers in both sets of markets exhaustedcredit near origination in anticipation of put option value, but this would leave open the questionof why origination growth was greater in the treated Sand State markets, particularly among loanswith small initial draws.Two natural identification concerns relate to sample selection and liquidity demand. In terms ofselection, borrowers with credit limits greater than the value of their homes have little incentiveto leave their homes. We might think that exiting borrowers would have less interest in exploitingthe embedded put options and might have experienced less severe home price declines. However,we find no difference in home price declines between terminating and surviving loans, consistentwith terminations arising primarily through death or severe health shocks. The larger price declinesin metropolitan areas where put options generally expire in the money could have led to reducedliquidity relative to the comparison metropolitan areas. In the latter areas, roughly a third of ter-minations were refinances, so some borrowers saw borrowing limits rise. However, we do not findevidence that borrowers use credit more when their homes decline in value less in the comparisongroup.Among the borrowers using little credit in the first year of the loan’s life, only a small minority ex-haust credit before termination, even in markets where loan-to-value ratios at termination averagedroughly 150%. The stakes were high for loans originated near the cycle peak and terminated priorto the end of our sample period. Among terminated loans in the treated metropolitan areas on whichborrowers withdrew less than 50% of available credit in the first year of the loan’s life, the mediangap between available credit and the greater of credit used or property value was approximately109$55,000. The mean “money left on the table” among all borrowers in the treated metropolitan ar-eas whose loans originated near the peak and terminated prior to the end of our sample was over$18,000.The apparent lack of ruthlessness among borrowers terminating with in the money put options sug-gests that something other than borrowers’ calculation of put option value drove adverse selection.The Sand States saw larger price booms than our comparison markets, but the difference is smallrelative to the difference between appreciation in the control markets and the rest of the US. De-spite this, the ratio in origination growth from a baseline prior to the price boom between the SandStates and the control markets was 6.4 to 2.3, whereas growth was greater outside both groups ofmarkets (2.4) than in the control set of metropolitan areas. Given their historically elastic supplyand low real price growth, borrowers in the treated markets may not have anticipated the capitalgains they saw, such that a modified Artle and Varaiya (1978) consumption smoothing story maybe a satisfactory explanation for the relative surge in the Sand States.Lenders may have played an important role in the spatial distribution of both price crashes andHECM originations. We leave empirical verification of that conjecture to future research. Thefact that loans originated through the failed lender, Financial Freedom, feature greater credit use isinteresting. However, Financial Freedom had a slightly smaller market share in the extreme crashSand State markets than in the comparison markets. The prevalence of retirees in the Sand Statesmight explain a greater level of originations if there are economies of scale in lending. This wouldnot readily explain the difference in growth rates of originations between the periods before andduring the home price peak, though: the relative rate of growth in retirement population over thisperiod must have been small compared to the relative growth of HECM loans.3636Given the specificity of reverse mortgages, their generally high growth through time, and the appeal of originatingforward mortgages through the cycle peak, it could have taken until the mid-2000s for reverse mortgage originators toconcentrate on these markets.110A natural direction for future research would be to distinguish among explanations for borrowers’failure to “ruthlessly” maximize home equity under HECM rules. Natural candidate explanationsinclude: a lack of understanding of HECM incentives, fear of credit damage, reluctance to behaveunethically, Medicaid incentives, and sudden incapacity. Unlike in the forward market, amongborrowers terminating their loans, future call option value on the home cannot explain a failure toexploit the put option. The structure of incentives also rule out precautionary savings or demand forbequests as compelling explanations. Borrowers have not had a very long time to adapt to decliningcollateral values, but some will likely never see gaps between collateral value and available creditas large as near the end of our sample period. Price increases much greater than accumulation ofloan balances and principal limits have reduced the moneyness of many implicit put options on themajority of in the money loans that remain active since the end of our panel in 2011.Fear of credit score or reputation damage from borrowing more than their home is likely to beworth at termination is worthy of exploration as a cause of non-ruthless credit use. Such fearswould not alter our main conclusions. As long as the expected value of these costs do not changebetween origination and termination, a potentially ruthless borrower only reluctant to draw creditabove collateral value for such reasons would presumably then not select into HECM for purposesof exploiting limited liability. Our approach would be invalid if borrowers selected into HECMwith ruthless intent, but then learned that exploiting limited liability were extremely costly. Basedon our discussions with industry professionals we do not see that as a plausible story. A relatedpossibility is that some borrowers who would use HECM for liquidity purposes do not do so be-cause they do not want to borrow more than their homes will be worth. This would be a violation ofour assumption that liquidity demand is similar between the treated and comparison metropolitanareas. Conceivably, a large fraction of borrowers who exhaust credit to exploit the put option aremasked by an offsetting fraction of such “liquidity” borrowers reluctant to face the need for a shortsale.111Information may play some role. We find that among borrowers not exhausting near origination,the relationship between put option moneyness and credit exhaustion is slightly stronger amongloans that did not terminate through 2011 than among loans terminating 2009-2011. Borrowerswith loans active through our sample had longer to contemplate optimal credit use, consistent witha role for information. As FHA releases a longer panel of credit use, this conjecture should besubject to confirmation. The small difference in exhaustion sensitivity and the declining hazardinto exhaustion over loans’ lives that we find suggest this effect would be small.The near absence of ruthless credit use has mixed implications for the future of reverse mortgages.Credit use stands out as one dimension of borrower behavior in which incentive problems have notarisen as a problem in HECM design. Loan durations have been longer in worse performing loans,which is a problem when insurance fees are charged periodically rather than all up-front. Liquidityconstrained, and perhaps opportunistic, borrowers in markets experiencing large price declinesappear to have defaulted in some cases on property tax and insurance obligations, and results inCapone et al. (2010) suggest that maintenance on these properties has suffered, too.The fact that credit use is not highly responsive to put value calculation suggests that retaining aline of credit option in HECM, as FHA has recently done while eliminating other payment op-tions, would be desirable. If credit use were elastic with respect to put value, then charging intereston draws and allowing the loan balance to grow at the loan interest rate would almost guaran-tee insurance losses. With credit use apparently reflecting liquidity demand, the benefits of theflexibility inherent in a line of credit, relative to loans that require immediate lump sum, presum-ably outweigh any dynamic contracting cost. However, losses due to selection suggest that federalmortgage insurance pricing and limits should incorporate deviations from historical local price-rentratios.The results suggest that home price insurance is not an important determinant of HECM demand.112Borrowers may thus be reluctant to pay “high” fees and interest charges if they do not value theimplicit insurance these fees provide. Given a spread between the economic value of insurance andthat perceived by borrowers, loans with lower fees and smaller loan amounts, as in “HECM Saver”loans; or with mandatory annuities sufficient to cover interest, property taxes, and insurance, mayprovide a path to market growth.More generally, weak demand for HECM and an absence of ruthless put option behaviour suggestthat what appear to have been deviations from fundamental home values in the Sand States werenot viewed with skepticism by a large segment of the population. Older homeowners were able totake short positions in home prices, but available data on HECM credit use does not suggest thata large number saw value in such a short position. Adverse selection into HECM thus does notprovide a counter example to studies finding limited financial sophistication among the elderly orirrational home price expectations.The empirical analysis used in this paper captures the put option using a a series of year fixedeffects in a difference-in-difference framework. Further research efforts would extend the hazardrate framework used in the paper to explore the role that borrower type/location plays in credit uti-lization. With sufficient data on borrower characteristics, this approach would have the potential toprovide more detailed insights into the sources geographic variation in default and credit utilizationin HECM reverse mortgages.113Chapter 4Fear and Loathing of Oil Pipelines:Hunting for Disamenity Effects14.1 IntroductionPipelines are a source of political contention in North America. Opposition to pipelines suchas Dakota Access and Keystone XL in the US, and Energy East, Northern Gateway and TransMountain expansion projects in Canada has been both local and national. The greater part of theopposition has targeted the role of pipelines in abetting fossil fuel use and its effects on climatechange. Along pipeline paths, there has been significant local opposition motivated by concernover the environmental risk from pipeline spills. This paper uses a variety of static hedonic anddynamic event study methodologies to estimate the capitalization of this latter effect on residentialproperty values.2 Compared to existing work on the capitalization of environmental hazards, ouruse of a rich data set of transactions from a dense suburban environment along a 42 km stretchof pipeline in the Vancouver, BC Canada metropolitan area allows us to apply more precise anddetailed treatments of location relative to the pipeline alignment and account more completely forthe variety of externalities in the land use fabric than is the case in existing work. We find that thesimple parametric treatments of proximity in the existing literature are likely to suffer from bothmodel misspecification and omitted variable bias. Model misspecification may arise because thenegative proximity effects of environmental hazards may be highly localized. Omitted variable bias1This paper was co-authored with Tsur Somerville at the Sauder School of Business - Tsur.Somerville@sauder.ubc.ca2The data for this paper were collected as part of a consulting project for Kinder Morgan examining the effectsof oil pipelines on the values of nearby residential property values. This report is available from Canada’s Na-tional Energy Board as filing 2015-08-20 Trans Mountain Pipeline ULC B417-28 - Reply_Evidence-Appendix _9A-Landowner_Compensation - A4S7H5114may be a arise because environmental hazards are associated with forms of land use that imposeexternalities on nearby residential land uses that are independent of the hazard under study.Theother contribution of this paper is in parsing the effect of new information regarding a hazard,differentiating between a reminder of the risk imposed by the environmental hazard and a reminderof its presence.A consistent, reliable framework for assessing the effects of environmental externalities is neces-sary for appropriate cost benefit analysis on facility siting and compensation for spills, leaks, andother hazardous discharges. The existing literature on the effects of proximity to environmentalhazards on residential property values is highly varied and does not offer clear guidelines for as-sessing the magnitude of externalities. This is in part because the nature and awareness of hazardsdiffer dramatically between hazardous waste sites, high-voltage powerlines, landfills, leaking oilstorage tanks, and gas and oil pipelines, to mention a few. In general, work on pipelines tends tofind no effect of proximity on residential property values. However, new information about thereminder of risks as well as pipeline construction does in some studies result in lower values thanwhat would otherwise hold for nearby properties. We add to this literature and shed some light onthe patchwork of results by taking advantage of a far richer data set, both in terms of the volumetransactions, controls for externalities from a variety of other non-residential land uses, and the useof very fine-grained treatments for distance than in previous work.The results of this paper highlight the sensitivity of proximity effects to specification and control-ling for variation in land use types. We find that pipeline proximity results in lower property values,but only for the most immediately adjacent properties. Properties on the pipeline easement on aver-age are 5.5 percent lower in terms of value, while those adjacent to a property with an easement orone property further away, have 2.2 and 1.3 percent lower property values respectively.3 Properties3Properties within 100 meters and not on the easement have a 1.2 percent lower value. 80 percent of these are theproperties adjacent to or one further away from the easement.115further away are unaffected. However, these results are sensitive to the type of land use throughwhich a pipeline passes. The residential property adjacent to the the pipeline easement is 3.5 per-cent lower when the pipeline is located on a non-residential land use as compared with 1.6 percentlower when the land use is residential or open space; these figures can even function as a set ofgeneral controls for proximity to arterials and different land use types. In all forms, these effectsdecay extremely rapidly with distance, and are not evident more than two properties away from thepipeline easement. Compared to existing papers we find that estimates of the effect of proximity insimple parametric specifications for distance fall in magnitude when the model includes the effectsof other nearby land uses and once the model accounts for properties with an easement and thoseimmediately adjacent. This is consistent with the argument we make regarding specification andleft-out variable bias in tests for negative proximity effects.Using the same data and specification we also test for the effects of new information about risks oncapitalization. The particular contribution we make is differentiating between news that remindsbuyers of risks as compared with news that reminds them of the presence of a hazard. We con-duct two difference in difference event studies, one for a well-publicized localized spill along thepipeline in one of the communities studied herein the study area, and the second for the announce-ment by the pipeline’s owner of a proposal to twin the pipeline and nearly triple the pipeline’scapacity. We treat the former as a reminder of the risk associated with the pipeline and the latter asa reminder of the presence of the pipeline. In the six months following the spill, transaction pricesfor properties within 250 meters of the pipeline away from the spill site were 5 percent lower thanthose further away.4 However, this difference disappears by nine months. In contrast, just the re-minder that there is a pipeline is not enough to affect prices: there is no change by location relativeto the pipeline alignment in transaction prices following the expansion announcement.54The properties that experienced contamination did not sell so we did not estimate the direct effect of contamination.5Pipeline are buried and there presence is not necessarily known to those nearby. It is possible that the announcementhad no effect because buyers discounted the likelihood the proposed expansion would be allowed.116The contributions of this paper to the literature on the effect of environmental risks on house valuesin general and the effect of pipelines in particular lie in several areas. First, we present moredetailed and precise measures of proximity at a very localized level than are found in other papers.Second, we address the land use context of the pipeline easement itself and identify its criticalinfluence on the estimated effect of pipeline proximity on home value. Finally, using the same datawe take advantage of two different kinds of shocks to awareness of the pipeline’s presence and thepossible risks to see whether increase in both affects prices. In doing so we highlight the sensitivityof hedonic pricing of environmental hazards to specification bias because of parametric treatmentof proximity and omitted variable bias because these hazards are associated with land uses thatthemselves are disamenities for residential properties.The paper follows the standard framework. Immediately below is a brief review of the existingliterature on the capitalization of environmental risks, primarily for oil pipelines, on house prices.This is followed by a description of the oil pipeline in question, the geography of proximity, and thetransaction data we use in this paper. Finally we present the empirical tests of the effects of pipelineproximity and information shocks, the spill, and expansion announcement, on house prices.4.2 Literature ReviewThere is an extant literature that explores the effects of proximity to oil and gas pipelines on houseprices. This research is part of the more general literature on the negative externalities of environ-mental hazards that measures the size of these effects through the relationship between exposureintensity in geographic space and the prices of residential real estate. This literature covers a verybroad range of work on environmental externalities. Surveys of this literature include review papersby Farber (1998), Boyle and Kiel (2001), Jackson (2001), Braden et al. (2011) and Sigman andStafford (2011). Their reviews cover papers on the effects on quality of life and risks to personsand property from a wide range of undesirable land uses (e.g. hazardous waste sites, power lines,117landfills, incinerators, and pipelines) as measured through a hedonic house price equation with theinclusion of a measure of proximity to the hazard as a right hand side covariate. The methodolo-gies range from simple static hedonic pricing equations to event studies or difference in differencesapproaches. With the latter, the natural experiment is either some new information about the riskof the hazard or change in its status, such as approval, construction, start of operation, closure,finding of hazard, or remediation. Overall, it is hard to draw specific conclusions about the natureof proximity due to the varied nature of the externalities associated with each of the different typesof environmental hazard studied in these works and the wide variation in the degree to which pa-pers address surrounding land uses, the distances over which effects are estimated, and the problemof non-random location of hazardous sites. For example, the findings of the hazardous waste siteliterature suggest that house prices fall with proximity to the noxious location, but not always asthe effects are sensitive to other nearby land uses and neighbourhood features.Within this broad group of work, two streams of research more directly inform the research pre-sented here. The first examines the explicit effect of oil pipelines (and relevant effects of gaspipelines), on nearby properties. These papers primarily use a static hedonic analysis methodology,regressing property value against distance to the pipeline and a set of structure and lot characteris-tics. The second studies the economic impact of proximity following some news regarding the riskof the pipeline, either a spill on the pipeline under study or news about spills in general, both ofwhich might be expected to heighten awareness and increase the magnitude or the duration of theassociated price effects. This type of information shock or new information about the presence ofa pipeline is likely to be especially important for pipelines because the presence of pipelines is notnecessarily known to buyers. While environmental disamenities such as high voltage transmissionlines, industrial facilities, and landfills can be seen or have a sensory impact on nearby properties(e.g. air quality, noise, visual), pipelines do not, as they have little visual presence and, unless thereis a leak or spill, entail no ongoing harm to nearby properties. The concern with pipelines is the118risk of a catastrophic incident that results in loss of property value or complete loss of use due tocontamination, quality of life or, in the case of gas pipelines, injury or death because of an explo-sion. Since the risks posed by pipelines are limited to the risk of an accident, measuring the effectof proximity to oil pipelines on residential prices helps to parse the mix of impacts associated withdifferent types environmental hazards.The risk to property from a pipeline rupture is non-trivial, with partial losses from nearby oilcontamination on a property exceeding ten percent and for significant oil contamination a completeloss of use.6 The most consistent work in this area comes from studies that examine the impact ofcontamination arising from leaking from underground oil storage facilities. For example, Simonset al. (1999) find loss of value from contamination of nearby soils to be between 14 and 16 percent.Zabel and Guignet (2012) relate house prices to publicized and unpublicized sites with leakingstorage facilities. They find that only sites where the contamination is well known have proximityeffects in the absence of known contamination, where these negative effects can exceed 10 percent.There is a need to perhaps differentiate between oil and gas pipelines. The spills of the former canlead to contamination and loss of use, while the latter do not, with effects dissipating rapidly. Spillsand ruptures of the former do not represent an immediate risk of injury and death, while the risk ofan explosion is acute with the latter.In a survey on gas pipelines, Wilde et al. (2013) report no evidence of proximity price effects in theacademic and professional appraisal literature, either for proximity in general or in the aftermath ofruptures.7 Even a more extreme form of pipeline risk manifests little pricing risk. In Boxall et al.’s(2005) study of gas wells and pipelines in rural Alberta, the case of sour gas is examined. Sourgas is both more noxious and more dangerous if released than conventional natural gas, gasoline,6While property owners typically receive compensation for the loss in value, the value of a property may be impactedwell into the future as a result of ongoing stigma.7For example, Kinnard Jr et al.’s (1994) hedonic study on gas pipelines and Diskin et al.’s (2011) matched-pairappraisal of properties adjacent to gas pipeline right-of-ways in three Arizona suburban subdivisions, both fail to find anegative relationship between pipeline proximity and residential sales prices.119or crude oil.8 Pipelines appear to be associated with negative values, but this is likely a left-outvariable problem. Contingent on the presence of sour gas wells, the presence of pipelines doesnot result in a further erosion in value. Three general problems with this type of static analysisare: awareness of the presence of the hazard, and then methodological hazards may be located inlower land value locations, and they may also be associated with negative externality land uses forreasons that are distinct from the specific hazard.Using changes in information is common in other research on hazards, where hopefully the dif-ference in differences methodology reduces problems with left-out variable bias from a hazard’snon-random location and the effects of other unmeasured land uses. These studies examine either[1] changes in information regarding the presence of a hazard or [2] changes in the extent of the riskas environmental remediation occurs. For example, Dale et al. (1999) find that at a broad metrolevel, the price of houses near a shutdown lead smelter rose faster than elsewhere after both theclosure and completion of environmental remediation. McCluskey and Rausser (2003) measurelocal house prices following the decommissioning of a hazardous waste incinerator. They find thatthe negative impact on housing prices arising from proximity to the incinerator dissipated after theclosure, but only slowly. The inverse to these natural experiments is the siting and construction ofa new facility. Kiel and McClain (1995) demonstrate that the expected negative proximity priceeffects from a rumoured, then proposed, then constructed, and finally operated garbage incineratorwere only manifested during the construction phase. What is more, this discount partially dimin-ished after ongoing operations commenced. Together, these findings suggests that house pricesdo respond as expected to new information. However, the revelation of new information seems tooccur slowly, and is stronger during negative market conditions (Case et al., 2006). Studies of oil8Health and safety risks associated with sour gas facilities represent a special hazard regulations requiring minimumsetback distances between sour gas and oil facilities and residential land uses. In addition to setbacks, emergency planresponse zones (EPZs) are established around sour natural gas facilities, the size of these zones can range up to severalkilometers and the size is related to the maximum potential volumes or rates of release of gas. More more informationsee Boxall et al. (2005)120pipelines have used the effect of spills on un-contaminated properties in order to address the sameproblems with identification.Without visual clues to their presence, it is possible that the absence of any effects of pipelineson house values reflects information failures. Spills convey two types of information. First is thepresence of risk, by serving as a reminder of the pipeline’s existence and its potential to rupture.Pipelines are typically buried, their presence is not obvious with the exception of discrete markersidentifying the pipeline easement, so the spill is a reminder of the pipeline’s presence. Second,a spill conveys information on the magnitude of risk based on the severity of a spill event. Intotal then, a spill acts as a change in a potential buyer’s or seller’s information set. Consequently,researchers have used a spill along a given pipeline as a natural experiment to test for the effect ofpipeline proximity. If the event is used with a difference in differences approach, it can also addressthe bias problems mentioned above. One approach is to look at changes in the value of propertieslocated elsewhere along the pipeline easement. Simons (1999) find that following a spill on theColonial Pipeline in Fairfax County Virginia, the value of properties along the easement elsewherein the same county experienced a 4.3-5.5 percent drop. He does not, however, report whether thiseffect dissipates with time, and furthermore the number of transactions in the easement after thespill is quite small, 76 over a four year period. Similar results are reported for a pipeline spill thataffects local waterways, where in a small sample, short run price declines are as high as 11 percent(Simons et al., 2001). Related work that followed the impact of oil spills over time found a ratherquick dissipation of the negative effect on prices. Papers on the Deepwater Horizon oil spill offthe Gulf Coast find that the negative effects of the spill on coastal house prices were temporary:returning to normal within 101 days (Siegel et al., 2013) and no lasting statistically significanteffects on sales volume or price levels (Winkler and Gordon, 2013).Work by Hansen et al. (2006) finds that in the absence of an information shock, there are nonegative proximity effects from pipelines. Using data on properties close to a pipeline in a small121city in Washington, they find that the rupture and explosion of a gas pipeline has a negative effect onthe prices of residential properties nearby to the pipeline. Using a highly concave inverse distancespecification, they find that prior to the rupture and subsequent explosion on the Olympic Pipeline,there was no relationship between distance to either pipeline and house prices. However, followingthe explosion, which because of the tragic deaths of an adult and two children in a park adjacentto the rupture, was extremely well-known, properties 50 feet from the Olympic Pipeline were anestimated 5.5 percent lower than properties beyond 1,000 feet from the pipeline. Interestingly,there was no effect on the properties adjacent to the pipeline that did not experience a rupture andthus lacked apparent "stigma". The negative price effect reported was transient in its nature. Thediscount at which properties within 100 feet of the pipeline traded following the accident declinedby 18 percent between 6 months and a year after the event, and by 27 percent after two years.9What is striking, is that the proximity to a second pipeline that did not spill was unrelated to prices,both before and after the explosion on the Olympic Pipeline.The ability of participants in local property markets to understand the nuanced issue around pipelinerisk was highly apparent in the Hansen et al. (2006) study. However, their study does not explicitlymodel the mechanism through which new information flows. McCluskey and Rausser (2001) testfor the effect of the volume of local media information on the perception of risk and find an explicitrelationship, though they do not tie this to proximity. The same conjecture, that negative effectsdepend on information is explicitly tested by Freybote and Fruits (2015). They examine howproximity to a gas pipeline is affected by perception of risk. To measure risk they include a fixedeffect that takes on the value of one if a pipeline explosion with fatalities was reported in the news inthe same month as the transaction. They study a high pressure gas line over a 14-year period in ex-urban Oregon. Their findings suggest that property values are negatively correlated to the proximityof the property to the gas pipeline, and the effects was attenuated during the construction period and9It is worth noting that very few sales, about 110 occurred within 300 feet (approx. 90m) of the pipeline over themulti-year analysis.122after a pipeline failure resulted in death.10 An attractive feature of the data used in the paper is thatit covers transactions observed prior to the construction, during the construction, and afterwards.Interestingly, the negative effect found in the study appears only during construction, and not post-announcement and pre-construction, nor is it present following the commencement of operation.This suggests that either pipeline risk depends on observing the pipeline, or that the negative effectsthey observe are from the construction process and not the pipeline itself. The information effectcan work through other processes. Kask and Maani (1992) find that negative price effects frompipelines only occur during the construction phase, when the presence is apparent, but once thepipeline is buried and operational there is no effect.Two papers of note address issues in specification and left-out variable bias in ways that relatedirectly to our paper. In general the papers on proximity effects look at large areas and modeldistance with a very simple monotonic parametric measure. Even if there is a specification thatallows for non-parametric relationships, for instance using fixed effects for distance bands, theminimum distance are typically at least a quarter mile. An exception is François (2002) whomodels the effect on house prices of proximity to high voltage power lines with a high degree ofgranularity and allows for a highly flexible specification. He finds that price effects are sensitiveto distance, direction, and the extent of visual awareness in ways that are not explicitly linear.For many locations of environmental disamenities there can be other land uses that are consideredundesirable. Failure to account for them will result in a left-out variable problem, that biases thecoefficient on proximity away from zero. Taylor et al. (2016) correctly observe that sites withenvironmental hazards are also typically located near other land uses that may impose negativeexternalities on residential properties. They then account for this by including properties witha similar land use, but without a hazard in their empirical analysis. They find that the mix ofcommercial properties with negative effects on residential properties and sites with environmental10The functional form used in their regression specification means that prices do not change adjacent to the pipeline,but do at a distance of one mile: relative not absolute prices are affected by proximity.123risk have an additive effect on nearby residential land uses. Furthermore, they find that clean-upand remediation is not fully capitalized, and that a stigma effect remains.The existing literature suggests that in general there are no effects of proximity to a pipeline onhouse values. When a spill has occurred, units closer to a pipeline, even if they are not contami-nated, have lower values. There are a number of problems with these studies. First, pipeline workstudies a small geographic area or uses relatively few transactions, which in turn may be quiteheterogenous. Second, they do not adjust for the nature of the pipeline’s easement, where distancefrom green space can be expected to have a different effect on nearby properties than distance froman industrial area. Third, the treatment of distance is typically just a parametric continuous measureand imposes assumptions about the relationship between proximity and value. Our contribution tothis literature comes from having sufficiently rich data and events to address the problems identi-fied in the research for a single hazard. First, like François (2002) we test for price effects thatare not parametric in distance, allowing for highly granular effects at extremely close distances.Second, we account for a variety of other land uses that can affect property values whose locationsmay be correlated in space with the pipeline, as is the case in Taylor et al. (2016). In addition, weaccount for land uses through which the pipeline itself passes, trying to separate the pure pipelineeffect from its land use context. Finally, we have two information events, one a spill and one anannouncement. Comparing these two allows us to partially differentiate between the effects ofpresence and the type of risk.4.3 Data and MethodologyThis study uses data along a segment of the Trans Mountain Pipeline (TMPL) that traverses thecities of Burnaby, Coquitlam, and Surrey in Lower Mainland area of Vancouver, BC. Burnaby isthe pipeline terminus and the three cities reflect the western most and most urban section of thepipeline routing in British Columbia. The cities are all part of the Vancouver, BC Canada Census124Metropolitan Area (CMA), their combined population as of the 2011 census is 842,200, makingup 35 percent of the metro area’s 2.37m population, and 18.6 percent of the area’s land mass. Thepipeline was built during 1952-53 and runs 1,155 km from Edmonton, AB to Burnaby, BC. InBurnaby there is a tank farm for storage and a marine terminal for shipments. Initial capacity was150,000 barrels per day (bpd). Additional pump stations expanded capacity to 410,000 bpd by1973, with peak delivery of 381,871 bpd in 1972.11The data include all of the transactions between 2000 and 2013 of single-family detached proper-ties in a 1.0 km buffer along either side of the pipeline easement as it passes through three citiesin the Vancouver CMA. Figure 1 illustrates the area of study. The pipeline alignment primarilyruns through single-family residential neighbourhoods with stretches of industrial, commercial,and open space. The solid black line indicates the location of the easement of the Trans Moun-tain Pipeline. The location of the single-family detached property transactions in the sample arehighlighted in pink. As can be seen in Figure 4.1, the transactions cluster quite densely along thepipeline alignment.11From 2002 to 2010 throughput ranged from 200,000 to almost 300,000 bpd.125Figure 4.1: Study area with Transactions4.3.1 Identification challenges and proposed testsA clean test to measure the impact of pipeline proximity on house prices is challenging because theidentification is affected by both left out variable bias and endogeneity in the relationship betweenpipeline and residential location choices. Ideally a randomly located pipeline in a residential areawould allow for a straight-forward difference in differences test. In reality pipelines are not ran-domly located. The cost of constructing a pipeline is a function of its length, the cost of acquiringthe pipeline easement, and the degree of local opposition. Pipeline firms will tradeoff among thesefactors in determining a pipeline’s alignment.The identification issues in regressing pipeline proximity against house value can be broken into126four categories. First, the left out variable bias because pipelines may be located on inexpensiveland. Second, a more general hedonic regression problem of structure quality when houses builton less expensive land have lower unobserved quality. Third, the endogenous location of amenitiesand pipelines either because pipeline builders provide amenities as part of the approval process orbecause subsequently greenspace and other non-structural amenities are built on pipeline right ofways where permanent structures may not be constructed.The first case of left out variable bias is because of an unobserved factor that makes the land lessexpensive. If the pipeline was put through undesirable locations where land was cheap, then housesnear the pipeline will have lower prices because of the same poor location factor. A simple prox-imity regression would be biased by the unobserved factor and result in bias away from zero in thecoefficient on pipeline proximity, overestimating the effect of a pipeline on housing pricesThe second form of left out variable bias would have the opposite result. Here, unobserved unitcharacteristics, where lower quality arises from complementarity in land price and structure quality,higher quality homes are built on higher priced land. Simple structure attributes like floor areawill be captured in the hedonic regression, but unobserved quality will affect the coefficients ofcorrelated included attributes. If houses are lower quality because of lower land prices, eitherbecause of the excluded factor above or because proximity causes lower land prices, then this willbias the coefficient on proximity towards zero.Finally, there the measurement issues that arise because of endogeneity between pipeline andnearby land uses. We identify two possible mechanisms. First, if developers provide local ameni-ties as part of the approval process, then these amenities, assuming they are unobserved, offset thenegative effects of the pipeline leading us to underestimate the negative impact of the pipeline onhousing prices. In the second case, local governments turn pipeline easements into linear greenspace or create amenities on the space that does not involve permanent structures. To the extent127these are unobserved, then they would increase the value of nearby residences, biasing the coeffi-cient on the effect of pipelines themselves.The latter two effects are unlikely to bias our analysis. In our sample, endogeneity is not a seriousissue because most of the houses were built after the pipeline was in place. 92% of the propertiesin the sample were built after the pipeline was constructed, and only one property with an easementexisted prior to the pipeline construction. Thus, the pipeline presence would be an in-situ realityfor all residential development. More importantly, we control for the type of land use throughwhich the pipeline easement passes as well as whether properties are proximate to a variety ofnon-residential land uses and open space and parks. Our analysis should control for the presenceof any amenities associated with the easement.The absence of a random placement of the pipeline that would allow for a difference in differencestest means we cannot completely escape left-out variable issues. However, the very granular localland use controls let us take into account both local amenity and nearby noxious land uses, as wellas those that are the land use context of the pipeline easement. In addition, we have two differencein differences tests associated with two shocks: the first is a spill along the pipeline easement andthe second is the announced plans to build a second pipeline along the existing easement to triplecapacity. While not as complete a test as one with a random placement, it does offer some insighton the effects of proximity from two different types of random events that highlight risk (oil spill)and remind residents of the presence of the pipeline (expansion announcement).Because the area of study is located in densely populated urban landscape, our data addresses manyof the shortcomings of existing studies in many ways. First, the density of the urban area yields alarge volume of transactions. Second, in comparison to rural or ex-urban areas, the characteristicsof the single-family houses we study are relatively homogeneous.12 Finally, the quality of avail-12The pipeline transverses one other municipality in the metropolitan area, Langley. We exclude this stretch becausemuch of the area is agricultural with a quite heterogeneous mix of residential and agricultural land uses.128able geographical data is high. This is important because in a densely populated urban area, theinfluence of other nearby land uses make it challenging to identify and isolate the impact of thepipeline proximity. Thus we account for location relative to major and minor arterials, open space,and a wide range of different types of non-residential land uses. While not the primary focus of ourstudy, we go to great efforts to control for the influence of other types of nearby land uses that mayalso be influencing house prices in order address the concerns raised by Taylor et al. (2016).The large number of observations gives us flexibility to characterize the distance of a propertyfrom the pipeline using measures that more closely measure the properties which are in closeproximity to the pipeline. Proximity of a property to the pipeline is measured in three separateways: as a continuous function of distance (0 to 1.0 km), in discrete bands of distance away fromthe pipeline, and finally as an ordinal measure of adjacency. For the latter we rank each propertyby distance from the pipeline in the number of properties removed. A property with an easementwould have an adjacency of zero, a property adjacent to the property with an easement would havean adjacency of one, the next property, the value would be two. Adjacency is calculated along avector perpendicular to the pipeline easement. 13In addition to the challenge of measuring the distance from a pipeline, a further challenge is pro-vided by the fact that the pipeline easement land use itself is not constant over the alignment. Overour area of study, the pipeline easement occurs on residential, commercial, and industrial prop-erties; along or under major and minor arterial roads; through open or green space; and acrosscivic (government, institutional, and recreational) land uses. A property adjacent to the pipelineeasement that is green space might be affected differently than one that is adjacent to a pipelineeasement on an arterial road or a non-residential land use because of the amenity value of the green13The latter orders properties 0 to 3, then beyond, based on the number of properties that separate them from theeasement. This latter we think of as distance in information space, as information about the pipeline is highest for thoseclosest or with neighbours who have an easement. Distance in this case is in the flow of information across propertyowners, and measured by the number of properties between a landowner and the easement.129space. In our data, we are able to identify the pipeline land use context. We define the pipeline ease-ment land use context for a given property as the closest along a vector orthogonal to the pipeline.This allows us to test for the effect different pipeline land use contexts by interacting these pipelineland use contexts with proximity measures.4.3.2 Description of sample and summary statisticsOur data were obtained from Landcor Data Corporation, a commercial provider of housing datafor the province of British Columbia. The data includes detailed information on the terms andcharacteristics of every property that transacted in Burnaby, Coquitlam and Surrey between 2000and 2013. The property level information includes variables that describe the land and structuresuch as: lot size, floor area, number of bedrooms, the effective age, number of stories etc. Thetransaction information includes variables that describe the date and price of each property sale.For each property, we also obtain information on the distance to other land uses including: com-mercial, industrial; civic (government, institutional, and recreational); major and minor arterialroads; through open or green space. Distance to the pipeline is measured in three ways: continuousmeasure, distance bands, and adjacency measures.The sample in this paper is constructed to aid identification. To create the sample, we start with allof the single family properties which transacted between 2000 and 2013 in the cities of Burnaby,Surrey, and Coquitlam. The area of analysis is then restricted to the properties within a 1,000-meterband on either side of the pipeline easement. The choice of a 1,000m band is warranted by the needto the balance transaction volume against neighbourhood homogeneity. To mitigate the influenceof outliers, we winsorize the lot size, floor area and adjusted sales price at the 1% level. The finalsample consists of 12,419 observations. Summary statistics for the sample are reported in Table1304.1.Table 4.1: Descriptive statistics for all transactionsmean median sd min maxProperty Characteristics Variables:Log price 12.97 12.97 0.46 11.33 14.54Log of repeat sales index 4.21 4.25 0.29 3.52 4.61Repeat Sales index adjusted price 687,837 652,426 271,018 157,314 2,515,605Lot size (thousands of sq/ft) 8,263 7,649 3,297 3,709 46,174Floor area (thousands of sq/ft) 2,658 2,403 977 812 6,150Number of bedrooms 4.20 4.00 1.25 1.00 8.00Effective age of property 31.67 30.00 14.33 2.00 95.00Number of stories 1.42 1.00 0.49 1.00 2.00Single garage dummy variable 0.22 0.00 0.41 0.00 1.00Multi garage (Ordinal variable) 0.59 1.00 0.51 0.00 3.00Number of full bathrooms 2.03 2.00 1.17 1.00 6.00Number of partial bathrooms 0.88 1.00 0.73 0.00 6.00Dummy, =1 if detached with suite 0.25 0.00 0.44 0.00 1.00Geographic Control Variables:Dummy, =1 if property < 100m from civic land use 1 (park/golf course/open green space) 0.14 0.00 0.34 0.00 1.00Dummy, =1 if property < 100m from civic land use 2 (govt bldg/works yard/cemetary) 0.22 0.00 0.41 0.00 1.00Dummy, =1 if property < 100m from civic land use 3 (institutional land use) 0.26 0.00 0.44 0.00 1.00Dummy, =1 if property < 250m from industrial land use 0.06 0.00 0.24 0.00 1.00Dummy, =1 if property < 250m from commercial land use 0.21 0.00 0.41 0.00 1.00Dummy, =1 if property < 40m from major arterial road 0.07 0.00 0.25 0.00 1.00Dummy, =1 if property within 40m of minor arterial road 0.01 0.00 0.12 0.00 1.00Pipeline Proximity Variables:Distance to pipeline in km 0.47 0.46 0.30 0.00 1.00Observations 12,419The first group of variables reported in Table 4.1 detail the distribution of transaction prices andthe property characteristics that are the standard set of controls widely used in hedonic house pricestudies. The dependent variable in each of the following specifications is the log of the sales price.However, interpreting this can be challenging because the data spans thirteen years, a period overwhich house prices in metro Vancouver rose 143 percent. For purposes of exposition, we reportan adjusted house price measure in the descriptive statistics. This measure is the unadjusted pricedeflated to 2015 dollars using estimated city-specific house price indexes. The indexes use the Caseand Shiller (1989) version of the more general Bailey et al. (1963) repeat sales methodology. Theindex is constructed for each city using paired transactions of single family properties that lieoutside the pipeline corridor.In the baseline hedonic regression specification, the dependent variable is the unadjusted log price.131In addition to the individual property characteristics we include jurisdiction-quarter dummies, cen-sus tract dummies, and the city-specific repeat sales index as right hand side variables. The repeatsales index captures overall changes in prices in the jurisdictions, while the jurisdiction-quarterfixed effects capture variation at the city level in the pipeline corridor relative to the rest of eachcity. Similarly, the census tract fixed effects capture systematic variation at the neighbourhood levelalong the pipeline corridor.The second group of variables in Table 4.1 correspond to the constructed geographic control vari-ables. The controls include a set of dummy variables that take on the value of one if the propertylies within a specified distance of an alternative land use. These variables cover the various typesof civic, industrial, and commercial land uses observed in the sample. We group the civic landuses into three groups corresponding to their likely amenity impact. For example, civic land use 1includes land uses that are likely to be have a positive amenity value (e.g. parks, golf courses andopen green space). If the property is within 100m of any of these land uses, the indicator is equalto 1, otherwise the indicator is equal to 0. Similarly, civic land use 2 includes land uses that arelikely to be disamenities (e.g. public works yards, cemeteries, misc government buildings). Wealso include separate dummy for commercial and industrial land uses and proximity to major andminor arterial roads.Separately we compared the statistics in Table 4.1 with a subset of transacting properties within100m of the easement, about 13 percent of the sample. The primary difference is that mean lot sizefor properties close to the pipeline is 10 percent larger than the overall sample. This is principallybecause the mean lot size for properties with a pipeline easement is about 12,000 sq ft, which isabout 50 percent larger than the mean lot size in the overall sample. One reason for this is thatthe easement is 18m (59 ft) wide and construction or structures are not permitted on the ease-ment. properties without an easement are not notably different, so much of this mean differenceis likely an effect of the easement itself. House-floor area is similar across these groups, and the132characteristics for those adjacent are roughly similar.Finally, we turn our attention to the pipeline proximity variables. Panel (a) of Table 4.2 providesfrequency counts for both the ordinal ranking of properties from the pipeline easement. In thesample, a surprisingly large proportion of the properties are within three properties of the pipeline.Of these properties, 134 have a pipeline easement and 587 are 1 parcel from the pipeline.Panel (b) of Table 4.2 provides frequency counts for properties in each of the distance groups usedin our analysis. The excluded group is the 6,000 transactions between 500-1,000 meters from thepipeline.Table 4.2: Frequency counts on distance to pipeline easement and adjacencyPanel (a) : Pipeline Proximity - Adjacency Measures Count Proportion (%) Cumulative (%)Indicator : Pipeline easement on property 134 1.08 1.08Indicator : Property is 1 parcel from pipeline 587 4.73 5.81Indicator : Property is 2 parcels from pipeline 490 3.95 9.75Indicator : Property is 3 parcels from pipeline 421 3.39 13.14Total 1,498 13.14%Panel (b) : Pipeline Proximity - Distance Bands Count Proportion (%) Cumulative (%)Indicator : Property is 0 - 100m from pipeline 1,474 11.87 11.87Indicator : Property is 100 - 250m from pipeline 2,098 16.89 28.76Indicator : Property is 250 - 500m from pipeline 2,927 23.57 52.33Total 6,499 52.33%Not surprisingly, there is substantial overlap between the adjacency and distance band measuresreported in Table 4.2. To see how the two measures are related, we cross tabulate the distancebands and adjacency measures. As can be seen in Table 4.3, there is substantial overlap betweenproperties no more than one property removed from the easement property (adjacency values of 0,1, or 2) and being within 100m of the pipeline. Because the variables are so closely related, in the133regressions we test for the fixed effect of each of these variables separately.Table 4.3: Cross Tabulation of Adjacency and Distance Bands MeasuresProperty is Property is Property is1 Parcel from 2 Parcels from 3 Parcel from (%) TotalPipeline Pipeline Pipeline by RowProperty is 0 - 100m from pipeline 38.18 28.91 13.02 80.11Property is 100 - 250m from pipeline 1.00 3.00 14.95 18.96Property is 250 - 500m from pipeline 0.00 0.80 0.13 0.93(%) of Total by Column 39.19 32.71 28.1 100%Total Observations 1,4984.4 Regression Specification and ResultsThe statistical analysis uses the standard hedonic regression format. In Table 4.4 and subsequenttables, the natural log of the property price (P) is regressed against a set of structure and lot char-acteristics (Xc) with both linear and quadratic terms for floor area and lot size. Basic time andgeography factors are captured by census tract fixed effects and jurisdiction-specific year dummiesalong with the jurisdiction-specific quarterly house price index (XJ).14 We also include a set ofgeographic control variables (Xc) and pipeline proximity variables (XD). The hedonic regressionmodel specified above is defined in equation 4.1.ln(P) = α+βXc+δXJ+γXG+ψXD+ (4.1)In the simplest treatment in regression (1), lot and structure coefficients have all of the expectedsigns. In regression (2) we add distance to the pipeline to the baseline regression (1). In thiscase, distance is included as a simple parametric linear measure, as is typical in the literature on14The house price index is estimated from single family units in the jurisdictions that are not in the pipeline corridordescribed above.134Table 4.4: Baseline regression specifications with simple distance measuresDependent variable = ln(price) (1) (2) (3) (4) (5)Property CharacteristicsLot size (thousands of sq/ft) 0.0164*** 0.0167*** 0.0171*** 0.0175*** 0.0175***(0.0015) (0.0015) (0.0015) (0.0015) (0.0015)Lot size squared –0.0001*** –0.0001*** –0.0001*** –0.0001*** –0.0001***(0.0000) (0.0000) (0.0000) (0.0000) (0.0000)Floor area (thousands of sq/ft) 0.1566*** 0.1555*** 0.1507*** 0.1503*** 0.1503***(0.0100) (0.0101) (0.0099) (0.0099) (0.0099)Floor area squared –0.0073*** –0.0072*** –0.0069*** –0.0069*** –0.0069***(0.0015) (0.0015) (0.0015) (0.0015) (0.0015)Number of bedrooms –0.0053*** –0.0055*** –0.0057*** –0.0057*** –0.0057***(0.0018) (0.0018) (0.0018) (0.0018) (0.0018)Effective age of property –0.0024*** –0.0024*** –0.0024*** –0.0025*** –0.0025***(0.0002) (0.0002) (0.0002) (0.0002) (0.0002)Number of stories 0.0352*** 0.0350*** 0.0343*** 0.0346*** 0.0346***(0.0049) (0.0049) (0.0049) (0.0049) (0.0049)Single garage dummy variable 0.0071 0.0071 0.0061 0.0060 0.0060(0.0044) (0.0044) (0.0043) (0.0043) (0.0043)Multi garage (Ordinal variable) 0.0523*** 0.0526*** 0.0516*** 0.0509*** 0.0509***(0.0045) (0.0045) (0.0044) (0.0044) (0.0044)Number of full bathrooms 0.0028 0.0029 0.0042 0.0041 0.0041(0.0028) (0.0028) (0.0028) (0.0028) (0.0028)Number of partial bathrooms 0.0130*** 0.0133*** 0.0107*** 0.0105*** 0.0105***(0.0029) (0.0029) (0.0028) (0.0028) (0.0028)Dummy, =1 if detached with suite –0.0002 0.0001 0.0035 0.0035 0.0036(0.0041) (0.0041) (0.0040) (0.0040) (0.0040)Geographic Control VariablesDummy, =1 if property < 100m from civic land use 1 (park/golf course/open green space) 0.0110** 0.0110** 0.0110**(0.0054) (0.0054) (0.0054)Dummy, =1 if property < 100m from civic land use 2 (govt bldg/works yard/cemetary) –0.0261*** –0.0258*** –0.0258***(0.0051) (0.0051) (0.0051)Dummy, =1 if property < 100m from civic land use 3 (institutional land use) –0.0079** –0.0078** –0.0078**(0.0038) (0.0038) (0.0038)Dummy, =1 if property < 250m from industrial land use –0.0596*** –0.0600*** –0.0601***(0.0075) (0.0075) (0.0075)Dummy, =1 if property < 250m from commercial land use –0.0094** –0.0095** –0.0095**(0.0040) (0.0040) (0.0040)Dummy, =1 if property < 40m from major arterial road –0.1056*** –0.1062*** –0.1063***(0.0064) (0.0064) (0.0064)Dummy, =1 if property within 40m of minor arterial road –0.0400*** –0.0401*** –0.0401***(0.0125) (0.0125) (0.0125)Pipeline Proximity VariablesDistance to pipeline in km 0.0158** 0.0111* 0.0077 0.0077(0.0064) (0.0066) (0.0067) (0.0067)Pipeline Easement on Property –0.0515*** –0.0424(0.0151) (0.0259)Interaction : Easement Dummy = 1 x Lot Size –0.0007(0.0016)Census Tract Dummies Y Y Y Y YJurisdiction Quarter Dummies Y Y Y Y YAdj. R-square 0.867 0.867 0.871 0.871 0.871Number of Cases 12,419 12,419 12,419 12,419 12,419* p < 0.1, ** p < 0.05, *** p < 0.01. Standard errors in parentheses.135proximity effects. Here we find that properties one kilometer from the pipeline are 1.6 percent morevaluable (about $10,300 at the mean). Introducing fixed effects for other land uses in regression(3), proximity to major roads and a variety of civic, commercial, and industrial land uses, that forthe most part have the expected negative effects on nearby residential properties, lowers this effectof pipeline proximity to 1.1 percent and it remains marginally statistically different from zero. 15As per Taylor et al. (2016), this highlights the problems for any hedonic study on proximity toa negative amenity that does not carefully address other land uses. The largest effect, 10 percent,is for properties within 40 meters of a major arterial, which include the TransCanada, Barnett, orLougheed Highways, with the second largest negative effect, a 6 percent discount for propertieswithin 250 meters of an industrial land use.For regression (4) we include a dummy if a property has a pipeline easement, so the pipeline trans-verses the property. This results in a 5 percent lower price ($33,300 at the mean) and reduces thecoefficient on distance by one third. This suggest that distance effects observed in the literaturemay be highly localized and including a parametric continuous distance measure to a hazard forproperties that are a considerable distance away is a source of specification bias. Finally, in regres-sion (5) we test for variation in the easement effect. Specifically, we are interested in whether theeasement effect changes in relation to the size of the property. To do this we include an interactionbetween the easement dummy and the lot size. Although neither the coefficients for the easementof the interaction are statistically different from zero, the magnitude and sign of the coefficientssuggest a two part effect in which the per square foot discount for properties that have an easementis declining in lot size.Rather than rely on parametric relationships for proximity, we allow for a more general treatmentof distance to the pipeline and transaction price. In regression (1) of Table 4.5 we try four dis-15Within 100 meters of a civic land use, 250 meters of a commercial or industrial land use, and 40 meters from anarterial, with major and minor arterials treated separately. Civic land uses are type 1 - parks and schools, type 2 - dumpsand corporation yards, type 3 - municipal buildings and facilities.136tance bands: 0-100m, 100-250m, 250-500m, and the excluded 500m-1km, where the effect is fixedwithin a band. The second approach taken in regression (1) of Table 4.6 is to use ordinal distancefrom the pipeline, where 0 is a property with an easement, 1 is a property adjacent to the propertywith the easement, 2 is one further away, 3 one again, and 4 plus the excluded default. Both ap-proaches indicate that the effect of proximity is only for those properties closest to an easement andthose within 100 meters or those adjacent to an easement property. These results again suggest thatthe existing research that finds no effect is inaccurate because the treatment of proximity in thesepapers as a continuous variable over long distances is too coarse. This calls into question muchof the literature on pipelines that finds no effect of proximity, small effects (one to two percent ofvalue) are evident for very nearby properties. It is not surprising that in parametric specificationswith data stretching a mile or more from a pipeline, that these highly localized negative externali-ties would not be identified in the data.137Table 4.5: Distance in Discrete BandsDependent variable = ln(price) (1) (2)Pipeline Easement on Property –0.0547*** –0.0576***(0.0151) (0.0151)Dummy, =1 if property 0 - 100m from pipeline –0.0125**(0.0055)Dummy, =1 if property 100 - 250m from pipeline 0.0048(0.0049)Dummy, =1 if property 250 - 500m from pipeline 0.0037(0.0043)0 - 100m from pipeline x pipeline context = Civic/Comm/Ind/Utility –0.0416***(0.0115)0 - 100m from pipeline x pipeline context = Open/Residential/Res.Road –0.0104(0.0063)100 - 250m from pipeline x pipeline context = Civic/Comm/Ind/Utility –0.0054(0.0091)100 - 250m from pipeline x pipeline context = Open/Residential/Res.Road 0.0021(0.0057)250 - 500m from pipeline x pipeline context =Civic/Comm/Ind/Utility –0.0014(0.0073)250 - 500m from pipeline x pipeline context = Open/Residential/Res.Road 0.0018(0.0050)Property Control Variables Y YGeographic Control Variables Y YCensus Dummies Y YJurisdiction Quarter Dummies Y YAdj. R-square 0.871 0.871Number of Cases 12,419 12,419* p < 0.1, ** p < 0.05, *** p < 0.01. Standard errors in parentheses.Pipelines, unlike most other environmental risks, will be co-present with other land uses. The effecton a nearby property has to be a joint effect of the pipeline and its land use "context". We parsethe average effect across pipeline easement land use types by interacting the pipeline proximitymeasures with fixed effects for the easement land use type. The results are quite striking. It can bebeen seen in regression (2) of Table 4.5 that for properties within 100 meters of a pipeline, the onlynegative effect occurs when the pipeline passes under one of the land uses identified as negative inTable 4.4.138Table 4.6: Distance in Adjacency MeasuresDependent variable = ln(price) (1) (2)Pipeline Easement on Property –0.0568***–0.0567***(0.0149) (0.0149)Dummy, =1 if property 1 parcel from pipeline –0.0211***(0.0074)Dummy, =1 if property 2 parcels from pipeline –0.0143*(0.0079)Dummy, =1 if property 3 parcels from pipeline 0.0072(0.0085)1 parcel from pipeline x Pipeline context = Civic/Comm/Ind/Utility –0.0352**(0.0172)1 parcel from pipeline x Pipeline context = Open/Residential/Res.Road –0.0164*(0.0089)2 parcels from pipeline x Pipeline context = Civic/Comm/Ind/Utility –0.0673***(0.0159)2 parcels from pipeline x Pipeline context = Open/Residential/Res.Road 0.0000(0.0099)3 parcels from pipeline x Pipeline context = Civic/Comm/Ind/Utility –0.0127(0.0250)3 parcels from pipeline x Pipeline context = Open/Residential/Res.Road 0.0112(0.0101)Property Control Variables Y YGeographic Control Variables Y YCensus Dummies Y YJurisdiction Quarter Dummies Y YAdj. R-square 0.871 0.871Number of Cases 12,419 12,419* p < 0.1, ** p < 0.05, *** p < 0.01. Standard errors in parentheses.In this case the magnitude of the negative association rises almost fourfold. In regression (2)of Table 4.6 we observe that for the ordinal proximity measures, this effect is spread betweenthe properties immediately adjacent to the easement, and, those that are one property removed.While the increase is not as dramatic as with the pure distance measure, we again see highernegative proximity effects when the pipeline passes under less desirable land uses. These resultsare consistent with the work by Taylor et al. (2016) and offer a strong cautionary note to researchthat measures negative proximity effects for environmental hazards without controlling for jointlyco-located land uses.1394.5 Event Studies : Difference in Difference RegressionsUsing the same pipeline segment and data, we take advantage of two separate incidents that allowus to test whether new information regarding the presence of a pipeline or its environmental riskaffect nearby house prices. The first event we study is the "Westridge Oil Spill." The spill occurredon July 24, 2007 when a backhoe penetrated the Trans Mountain Pipeline spur in Burnaby, BritishColumbia. This stretch of pipeline runs a short distance from the Trans Mountain terminus onBurnaby Mountain to Port Metro Vancouver’s Westridge Marine Terminal. Although the spillonly released 1,500 barrels of heavy crude oil, the media attention that the spill received wasdisproportionately large. The second event of interest is the surprise announcement in May 2012by Kinder Morgan of their plans to nearly triple the pipeline’s capacity by twinning the pipelinealong its alignment. Kinder Morgan announced that the second pipeline will follow the existingalignment, but there are likely to be some changes to rationalize the route. It is unclear if thesecond pipeline will supplement or replace the existing 50-year-old pipeline. The two informationshocks allow us to examine the response to two types of information: a reminder of the risksassociated with proximity to pipelines and, second, a reminder of the presence of the pipeline withany associated reaction to its potential increase in size.For these tests we conduct simple difference in difference tests around each of the event dates. Theobjective of examining the divergence in short run price changes between houses closest to thepipeline alignment and those further away.Since there were fewer transactions in the narrower time window of analysis it was necessary tomake adjustments in the regression specification. We are not able to use measures for houses oneproperty away from the alignment or even within 100 meters, so we use 250 meters as the band forclose proximity.1404.5.1 Effect of SpillTable 4.7 below shows the difference in price appreciation for properties within 250 meters of thealignment compared to those that are 250 to 1,000 meters away after the spill. We use windowsfor the before and after element of the diff in diff methodology of 0-3 months, 0-6 months, and0-9, and then 0-12 months. Immediately after the spill - regression (1) - units within 250 meters ofthe alignment sold at a 6 percent discount compared with those further away relative to values inthe period previous to the spill. This suggests a rather large immediate reaction to the informationreminding residents of the risk associated with a spill, larger than the effects in the static analysisfrom above. Consistent with the idea of dissipation of awareness of the potential risk and concernabout the risk, this discount disappears with the longer windows.1616For comparison, we repeat the analysis of the spill using a larger band of 500 meters. The findings support twoclaims: the first that the valuation of spill risk declines with distance from the pipeline, as the absolute values of theestimated coefficient for distance band dummy, post-spill interaction are lower for the 0-500 meter band than the 0-250meter band in all time periods; and the second is re-confirmation that the estimation of this risk declines quickly withtime after the spill.141Table 4.7: Difference in Difference Regressions: Effect of Oil Spill (250m Bands)Dependent variable = ln(price) (1) (2) (3) (4)Dummy, =1 if property < 250m from pipeline 0.0250 0.0264 0.0183 0.0102(0.0227) (0.0178) (0.0149) (0.0125)Dummy, =1 if Sale 3 months post spill 0.0658**(0.0270)Sale < 3 months post spill x property < 250m from pipeline –0.0543*(0.0308)Dummy, =1 if Sale 6 months post spill 0.0671**(0.0269)Sale < 6 months post spill x property < 250m from pipeline –0.0517**(0.0234)Dummy, =1 if Sale 9 months post spill 0.0586**(0.0271)Sale < 9 months post spill x property < 250m from pipeline –0.0300(0.0201)Dummy, =1 if Sale 12 months post spill 0.0545**(0.0268)Sale < 12 months post spill x property < 250m from pipeline –0.0197(0.0167)Property Control Variables Y Y Y YGeographic Control Variables Y Y Y YCensus Tract Dummies Y Y Y YJurisdiction Quarter Dummies Y Y Y YAdj. R-square 0.767 0.743 0.736 0.745Number of Cases 570 922 1,286 1,735* p < 0.1, ** p < 0.05, *** p < 0.01. Standard errors in parentheses.Effect of spill robustness checkAs a simple robustness check we re-test the spill effects assuming that the spill occurs two yearsearlier, with a placebo spill in June 2005. The results in Table 4.8 show no evidence that the resultsfrom Table 4.7 occur because of some trend in the data that pre-dates the spill. All differencein difference interaction coefficients show relative difference before and after the placebo spillbetween units close to the pipeline and those further away.142Table 4.8: Robustness Check : Westridge Oil Spill Falsifications RegressionsDependent variable = ln(price) (1) (2) (3) (4)Dummy, =1 if property < 250m from pipeline 0.0002 –0.0058 0.0044 0.0058(0.0207) (0.0180) (0.0168) (0.0147)Dummy, =1 if Sale 3 months post spill –0.0069(0.0229)Sale < 3 months post spill x property < 250m from pipeline 0.0237(0.0277)Dummy, =1 if Sale 6 months post spill –0.0050(0.0222)Sale < 6 months post spill x property < 250m from pipeline 0.0202(0.0258)Dummy, =1 if Sale 9 months post spill 0.0162(0.0226)Sale < 9 months post spill x property < 250m from pipeline –0.0090(0.0227)Dummy, =1 if Sale 12 months post spill 0.0107(0.0216)Sale < 12 months post spill x property< 250m from pipeline 0.0040(0.0191)Property Control Variables Y Y Y YGeographic Control Variables Y Y Y YCensus Tract Dummies Y Y Y YJurisdiction Quarter Dummies Y Y Y YAdj. R-square 0.756 0.730 0.725 0.738Number of Cases 544 804 1,066 1,359* p < 0.1, ** p < 0.05, *** p < 0.01. Standard errors in parentheses.4.5.2 Effect of Expansion AnnouncementOur second natural experiment uses the pipeline expansion announcement. There is no coveragein the local media prior to the May 2012 announcement of a possible expansion, so we treat itas a surprise event. The information content in the event includes a reminder of the presenceof the pipeline, and possibly heightened risk at an unknown future date from the probability ofthe expansion approval. We apply the standard difference in difference methodology around theannouncement date.143Table 4.9: Pipeline Expansion Announcement Event Study RegressionsDependent variable = ln(price) (1) (2) (3) (4)Dummy, =1 if property < 250m from pipeline –0.0283 –0.0274 –0.0333** –0.0228**(0.0241) (0.0186) (0.0147) (0.0111)Dummy, =1 if Sale < 3 months post announcement 0.0095(0.0197)Sale < 3 months post announcement x property < 250m from pipeline –0.0104(0.0368)Dummy, =1 if Sale < 6 months post announcement 0.0132(0.0181)Sale < 6 months post announcement x property < 250m from pipeline –0.0158(0.0257)Dummy, =1 if Sale < 9 months post announcement 0.0117(0.0176)Sale < 9 months post announcement x property < 250m from pipeline –0.0094(0.0211)Dummy, =1 if Sale < 12 months post announcement 0.0096(0.0171)Sale < 12 months post announcement x property < 250m from pipeline –0.0068(0.0166)Property Control Variables Y Y Y YGeographic Control Variables Y Y Y YCensus Tract Dummies Y Y Y YJurisdiction Quarter Dummies Y Y Y YAdj. R-square 0.815 0.826 0.844 0.869Number of Cases 401 603 808 1,174* p < 0.1, ** p < 0.05, *** p < 0.01. Standard errors in parentheses.As can be seen in Table 4.9, we do not observe a differential price response following the expansionannouncement between properties within 250 meters of the pipeline easement and properties 250-1,000m from the easement.This suggests either full awareness of the pipeline presence for those properties affected by it,which above we show to be a very small highly localized group, and that the announcement of anexpansion was treated by the buyers or sellers as an indication of higher risk.This is consistent with Freybote and Fruits (2015), where they only observe pipeline proximityeffects during construction. Our interpretation of the results in Table 4.9 combined with our earlierfindings, is that except in the case of an event like a spill that heightens subjective assessments of144risk, only those properties closest to a pipeline experience a proximity discount. This is not due toa lack of market awareness, but for the reason that properties not within one to two properties ofthe alignment, the presence is not treated as a risk.4.6 Summary and ConclusionsThis paper explores the existence of negative externalities from oil pipelines on nearby residencesby testing for the effect of proximity to oil pipelines on residential prices. Our results help resolvesome of the inconsistent findings in existing research in this area. There is also a more generalcontribution: first, showing how highly localized negative proximity effects can be. In our datait is only the closest one or two properties that experience negative effects on value from nearbypipelines. Second, as with Taylor et al. (2016), we observe that the effects of proximity to aparticular environmental nuisance, in our case the pipeline alignment, are mixed together with ef-fects from other negative externality land uses that are spatially correlated with the environmentalnuisance. For example, when the pipeline transverses industrial and commercial land uses, the neg-ative proximity effect is at least twice as large as when the pipeline is passing through a residentialarea. These immediate effects exist even though we have geographic controls for more generalproximity to non-residential land uses and major and minor arterial roads.Unlike the existing literature we find that properties close to an oil pipeline do experience a declinein value, but this only holds for the closest one or two properties, which are all within 100 metersof the pipeline. A residential property with an easement has a 5.7% lower value, while a propertyadjacent to such a property has a price that is 2.1% lower than a more distant unit. While a propertythat is one more property away has a 1.4% lower sale price, the effects are lower in absolutemagnitude and not statistically different than zero further away. Given that these narrow effects areonly for the closest properties, it is not surprising that the literature using parametric specificationshas not previously found a relationship between negative values and closeness to an oil pipeline145except following a spill. Interacting these proximity by the type of land use further strengthensour contention of the importance in modelling proximity and land use effects with fine granularity.When the pipeline passes through non-residential land uses, the negative proximity effects aremuch higher than when the easement is on residential or open space land uses. In the latter case,the negative effect for an adjacent property falls from 2.1% lower to 1.6% lower.Our difference in difference analysis supports the findings elsewhere that risk assessments are af-fected by recent pertinent information. Interestingly, our results indicate that this affect is transient.Within nine months of a spill, the heightened negative proximity effects disappear elsewhere alongthe pipeline. As with our static price effects, these only apply to closer units, but they extend onaverage throughout a 250-meter band, instead of a 100-meter band, and have an average negativeeffect of 5.2% instead of the 100-meter band 1.2% lower price in the absence of a spill. The tests onpipeline expansion announcement yield no difference in difference results. In combination thesefindings suggest that it is heightened risk and not a renewed reminder of the presence of the pipelinethat causes lower transaction prices. Pipelines do affect property values, but the effect is narrowand focused. Our data also suggest that parametric continuous distance treatments of proximityfound in much of the literature are a very poor way to test for negative spatial externalities fromenvironmental hazards.This paper sheds light on some of the factors that contribute to large variations in results of studiesthat examine the effect of proximity to environmental hazards on residential prices. The richness ofour data permits us to model distance from the hazard more finely and account for a broad range ofother land uses, proximity to which can be expected to affect residential property values. We findthat both have notable effects on the relationship between proximity and house prices. The moreprecise tests here show that proximity may only matter at very short distances, and that the failureto account for other land uses, and even the land use type of the hazard, will bias estimates awayfrom zero.146Chapter 5ConclusionsThis thesis is a collection of three essays in Real Estate Finance. Since the essays are presentedin self-contained chapters, discussion of the contribution and position of each essay in the broaderfinance literature is provided in the conclusion specific to each paper.147References2016 canadian cmbs market outlook, 2016. URL ExtractedAugust 2017.Sumit Agarwal, Brent W Ambrose, Souphala Chomsisengphet, and Chunlin Liu. The role of softinformation in a dynamic contract setting: Evidence from the home equity credit market.Journal of Money, Credit and Banking, 43(4):633–655, 2011.George A. Akerlof. The market for" lemons": Quality uncertainty and the market mechanism.The Quarterly Journal of Economics, 84(3):488–500, 1970.B Ambrose, J Benjamin, and P Chinloy. Bank and Nonbank Lenders and the CommercialMortgage Market. Journal of Real Estate Finance and Economics, 26(1):81–94, 2003.Brent Ambrose, Michael Shafer, and Yildiray Yildirim. The impact of tenant diversification onspreads and default rates for mortgages on retail properties. The Journal of Real Estate Financeand Economics, (Forthcoming):1–32, 2016. URL W. Ambrose and Anthony B. Sanders. Commercial mortgage-backed securities:Prepayment and default. The Journal of Real Estate Finance and Economics, 26(2):179–196,2003.Xudong An. Macroeconomic conditions, systematic risk factors, and the time series dynamics ofcommercial mortgage credit risk. University of Southern California Ph. D. PhD thesis,Dissertation, 2007.Xudong An and Anthony B Sanders. Default of commercial mortgage loans during the financialcrisis. In 46th Annual AREUEA Conference Paper, 2010.Xudong An, Yongheng Deng, and Stuart A Gabriel. Asymmetric information, adverse selection,and the pricing of cmbs. Journal of Financial Economics, 100(2):304–325, 2011.Xudong An, Yongheng Deng, Joseph B Nichols, and Anthony B Sanders. What is subordinationabout? credit risk and subordination levels in commercial mortgage-backed securities (cmbs).The Journal of Real Estate Finance and Economics, 51(2):231–253, 2015.148Wayne R Archer, Peter J Elmer, David M Harrison, and David C Ling. Determinants ofmultifamily mortgage default. Real estate economics, 30(3):445–473, 2002.Roland Artle and Pravin Varaiya. Life cycle consumption and homeownership. Journal ofEconomic Theory, 18(1):38–58, 1978.Martin J. Bailey, Richard F. Muth, and Hugh O. Nourse. A regression method for real estate priceindex construction. Journal of the American Statistical Association, 58(304):933–942, 1963.Ada-Helen Bayer and Leon Harper. Fixing to stay: a national survey of housing and homemodification issues. Research report, AARP, 2000.Sreedhar T Bharath, Sandeep Dahiya, Anthony Saunders, and Anand Srinivasan. Lendingrelationships and loan contract terms. Review of Financial Studies, 24(4):1141–1203, 2011.Peter C Boxall, Wing H Chan, and Melville L McMillan. The impact of oil and natural gasfacilities on rural residential property values: a spatial hedonic analysis. Resource and energyeconomics, 27(3):248–269, 2005.Melissa Boyle and Katherine Kiel. A survey of house price hedonic studies of the impact ofenvironmental externalities. Journal of real estate literature, 9(2):117–144, 2001.John B Braden, Xia Feng, and DooHwan Won. Waste sites and property values: a meta-analysis.Environmental and Resource Economics, 50(2):175–201, 2011.Michael Bradley and Michael R Roberts. The structure and pricing of corporate debt covenants.The Quarterly Journal of Finance, 5(02):155–183, 2015.Jeffrey R. Brown. Rational and behavioral perspectives on the role of annuities in retirementplanning. Working Paper 13537, NBER, 2007.Jeffrey R. Brown, Jeffrey R. Kling, Sendhil Mullainathan, and Marian V. Wrobel. Why don’tpeople insure late-life consumption? A framing explanation of the under-annuitization puzzle.American Economic Review, 98(2):304–09, May 2008.Consumer Financial Protection Bureau. Reverse mortgages. Report to congress, U.S.Government, 2012.Andrew Caplin. Turning assets into cash: Problems and prospects in the reverse mortgageindustry. In Olivia S. Mitchell, Zvi Bodie, P. Brett Hammond, and Stephen Zeldes, editors,Innovations in Retirement Financing, page chapter 11. University of Pennsylvania Press,Philadelphia, 2002.Andrew Caplin, Sewin Chan, Charles Freeman, and Joseph Tracy. Housing Partnerships: A newapproach to markets at a crossroads. MIT Press, Cambridge, 1997.149Charles A. Capone, Karen L. Chang, and Colin A. Cushman. Identification of home maintenancerisk in reverse mortgages: An empirical examination of home price appreciation among HECMborrowers. working paper, U.S. Department of Housing and Urban Development, 2010.Bradford Case, Peter F Colwell, Chris Leishman, and Craig Watkins. The impact ofenvironmental contamination on condo prices: A hybrid repeat-sale/hedonic approach. RealEstate Economics, 34(1):77–107, 2006.Karl E. Case and Robert J. Shiller. The efficiency of the market for single-family homes. TheAmerican Economic Review, 79(1):125–137, March 1989.John Chalmers and Jonathan Reuter. How do retirees value life annuities? Evidence from publicemployees. Review of Financial Studies, 25(8):2601–2634, 2012.Brian A Ciochetti, Yongheng Deng, Bin Gao, and Rui Yao. The termination of mortgage contractsthrough prepayment and default in the commercial mortgage markets: a proportional hazardapproach with competing risks. Real Estate Economics, 30(4):595–633, 2002.Andrew Cohen and Mark D Manuszak. Ratings competition in the cmbs market. Journal ofMoney, Credit and Banking, 45(s1):93–119, 2013.Larry Dale, James C Murdoch, Mark A Thayer, and Paul A Waddell. Do property values reboundfrom environmental stigmas? evidence from dallas. Land Economics, pages 311–326, 1999.Thomas Davidoff. Housing, health, and annuities. Journal of Risk and Insurance, 76(1):31–52,2009.Thomas Davidoff. Home equity commitment and long-term care insurance demand. Journal ofPublic Economics, 94(1-2):44–49, 2010.Thomas Davidoff. Can ‘high costs’ justify weak demand for the Home Equity ConversionMortgage? Working paper, Sauder School of Business, University of British Columbia, 2012.Thomas Davidoff. Supply elasticity and the housing cycle of the 2000s. Real Estate Economics,41(4):793–813, 2013.Thomas Davidoff. Reverse mortgage demographics and collateral performance. Working paper,University of British Columbia, 2014.Thomas Davidoff and Gerd Welke. Selection and moral hazard in the reverse mortgage market.working paper, UC Berkeley, 2006.Thomas Davidoff and Gerd Welke. Hecm refinance. working paper, University of BritishColumbia, 2013.150Yongheng Deng and John M. Quigley. Woodhead behavior and the pricing of residentialmortgages. Working Paper Series W00-004, UC Berkeley Program On Housing and UrbanPolicy, 2004.Yongheng Deng, John Quigley, and Robert Van Order. Mortgage terminations, heterogeneity andthe exercise of mortgage options. Econometrica, 68(2):275–307, March 2000.Yongheng Deng, Erik Devos, Shofiqur Rahman, and Desmond Tsang. The role of debt covenantsin the investment grade bond market–the reit experiment. The Journal of Real Estate Financeand Economics, pages 1–21, 2015.Ilia D Dichev and Douglas J Skinner. Large–sample evidence on the debt covenant hypothesis.Journal of accounting research, 40(4):1091–1123, 2002.Barry A Diskin, Jack P Friedman, Spero C Peppas, and Stephanie R Peppas. The effect of naturalgas pipelines on residential value. Right of Way, pages 24–27, 2011.Stephen Farber. Undesirable facilities and property values: a summary of empirical studies.Ecological Economics, 24(1):1–14, 1998.Amy Finkelstein and James Poterba. Adverse selection in insurance markets: Policyholderevidence from the u.k. annuities market. Journal of Political Economy, 112(1):183–208, 2004.Christopher L. Foote, Kristopher S. Gerardi, and Paul S. Willen. Why did so many people makeso many ex post bad decisions? The causes of the foreclosure crisis. Technical Report 12-2,Federal Reserve Bank of Boston, 2012.Charles Foster and Robert Van Order. An option-based model of mortgage default. HousingFinance Review, 3(4):351–372, 1984.Rosiers François. Power lines, visual encumbrance and house values: a microspatial approach toimpact measurement. Journal of Real estate research, 23(3):275–302, 2002.Julia Freybote and Eric Fruits. Perceived environmental risk, media, and residential sales prices.Journal of Real Estate Research, 37(2):217–244, 2015.Charles Gamm and Christopher Kane. Canadian cmbs: “low hanging fruit” for u.s. fixed incomeinvestors. Technical Report 3, CRE Finance Council, 2013.Edward L. Glaeser, Joseph Gyourko, and Albert Saiz. Housing supply and housing bubbles.Journal of Urban Economics, 64(2):198–217, 2008.William H Greene. Econometric analysis, harlow, 2012.R.A. Grovenstein, J.P. Harding, CF Sirmans, S. Thebpanya, and G.K. Turnbull. Commercial151mortgage underwriting: How well do lenders manage the risks? Journal of HousingEconomics, 14(4):355–383, 2005.Julia L Hansen, Earl D Benson, and Daniel A Hagen. Environmental hazards and residentialproperty values: Evidence from a major pipeline event. Land Economics, 82(4):529–541, 2006.Donald Haurin, Chao Ma, Stephanie Moulton, Maximilian Schmeiser, Jason Seligman, and WeiShi. Spatial variation in reverse mortgages usage: House price dynamics and consumerselection. The Journal of Real Estate Finance and Economics, pages 1–26, 2014. ISSN0895-5638.Christoph Herpfer. The role of bankers in the us syndicated loan market. Swiss Finance InstituteLausanne Working Paper, pages 1–59, 2017.Victoria Ivashina and David Scharfstein. Bank lending during the financial crisis of 2008. Journalof Financial economics, 97(3):319–338, 2010.Thomas Jackson. The effects of environmental contamination on real estate: A literature review.Journal of Real Estate Literature, 9(2):91–116, 2001.Susan B Kask and Sholeh A Maani. Uncertainty, information, and hedonic pricing. LandEconomics, pages 170–184, 1992.Benjamin J. Keys, Tanmoy Mukherjee, Amit Seru, and Vikrant Vig. Did securitization lead to laxscreening? Evidence from subprime loans. Quarterly Journal of Economics, 125(1):307–362,2010.Katherine A Kiel and Katherine T McClain. House prices during siting decision stages: the caseof an incinerator from rumor through operation. Journal of Environmental Economics andManagement, 28(2):241–255, 1995.William N Kinnard Jr, Sue Ann Dickey, and Mary Beth Geckler. Natural gas pipeline impact onresidential property values: An empirical study of two market areas. Right of Way, pages 26–9,1994.Paul Krugman. That hissing sound. New York Times, 2005. Op-Ed, August 8.Deborah Lucas and Damien Moore. The student loan consolidation option: An analysis of anexotic financial derivative. Working Paper 2007-5, Congressional Budget Office, 2007.Annamaria Lusardi, Olivia S. Mitchell, and Vilsa Curto. Financial literacy and financialsophistication among older Americans. Working Paper 15469, NBER, 2009.Brian A Maris and William Segal. Analysis of yield spreads on commercial mortgage-backedsecurities. The Journal of Real Estate Research, 23(3):235, 2002.152Christopher Mayer, Tomasz Piskorski, and Alexei Tchistyi. The inefficiency of refinancing: Whyprepayment penalties are good for risky borrowers. Journal of Financial Economics,forthcoming, 2012.Christopher J. Mayer, Edward Morrison, Tomasz Piskorski, and Arpit Gupta. Mortgagemodification and strategic behavior: Evidence from a legal settlement with countrywide.Working Paper 17065, National Bureau of Economic Research, 2011.Jill J McCluskey and Gordon C Rausser. Estimation of perceived risk and its effect on propertyvalues. Land Economics, 77(1):42–55, 2001.Jill J McCluskey and Gordon C Rausser. Hazardous waste sites and housing appreciation rates.Journal of Environmental Economics and Management, 45(2):166–176, 2003.Thomas Miceli and C.F. Sirmans. Reverse mortgage and borrower maintenance risk. Journal ofthe American Real Estate and Urban Economics Association, 22(2):257–299, 1994.Makota Nakajima and Irina A. Telyukova. Reverse mortgage loans: A quantitative analysis.Working paper, University of California, San Diego, 2013.Frank Nothaft and JL Freund. The evolution of securitization in multifamily mortgage marketsand its effect on lending rates. Journal of Real Estate Research, 25(2), 2003.Tomas J. Philipson and Gary S. Becker. Old-age longevity and mortality-contingent claims.Journal of Political Economy, 106(3), 1998.Donald L. Redfoot, Ken Scholen, and S. Kathi Brown. Reverse mortgages: Niche product ormainstream solution? Report on the 2006 AARP National Survey of Reverse MortgageShoppers. Report 2007-22, AARP, 2007.Philip A Seagraves and Jonathan A Wiley. The cap rate spread: A new metric for commercialunderwriting. Real Estate Economics, 44(2):490–520, 2015.Hui Shan. Reversing the trend: The recent expansion of the reverse mortgage market. Real EstateEconomics, 39(4):743–768, 2011.Robert Shiller and Allan Weiss. Moral hazard in home equity conversion. Real Estate Economics,28(1):1–31, 2000.Robert J. Shiller. Macro Markets. Clarendon Press, Oxford, 1993.Christy Siegel, Steven B Caudill, and Franklin G Mixon. Clear skies, dark waters: The gulf oilspill and the price of coastal. Economics and Business Letters, 2(2):42–53, 2013.Hilary Sigman and Sarah Stafford. Management of hazardous waste and contaminated land.Annu. Rev. Resour. Econ., 3(1):255–275, 2011.153Robert A Simons. The effect of pipeline ruptures on noncontaminated residentialeasement-holding property in fairfax county. Appraisal Journal, 67:255–263, 1999.Robert A Simons, William M Bowen, and Arthur J Sementelli. The price and liquidity effects ofust leaks from gas stations on adjacent contaminated property. Appraisal Journal, 67:186–194,1999.Robert A Simons, Kimberly Winson-Ceideman, and A Brian. The effects of an oil pipelinerupture on single-family house prices. 2001.Jonathan S. Skinner. Is housing wealth a sideshow? In David A. Wise, editor, Advances in theEconomics of Aging, pages 241–271. NBER and University of Chicago Press, Chicago, 1996.Clifford W Smith and Jerold B Warner. On financial contracting: An analysis of bond covenants.Journal of financial economics, 7(2):117–161, 1979.Richard Stanton and Nancy Wallace. Cmbs subordination, ratings inflation, and regulatory-capitalarbitrage. UC Berkeley: Fisher Center for Real Estate and Urban Economics Working Paper,December 2012.Joseph E Stiglitz and Andrew Weiss. Credit rationing in markets with imperfect information. TheAmerican economic review, 71(3):393–410, 1981.Laura O Taylor, Daniel J Phaneuf, and Xiangping Liu. Disentangling property value impacts ofenvironmental contamination from locally undesirable land uses: Implications for measuringpost-cleanup stigma. Journal of Urban Economics, 93:85–98, 2016.S Titman, S Tompaidis, and S Tsyplakov. Determinants of Credit Spreads in CommercialMortgages. Real Estate Economics, 33(4):711–738, 2005.Sheridan Titman and Sergey Tsyplakov. Originator performance, cmbs structures, and the risk ofcommercial mortgages. Review of Financial Studies, 23(9):3558–3594, 2010.Kerry D. Vandell. How ruthless is mortgage default? A review and synthesis of the evidence.Journal of Housing Research, 6(2):245–264, 1995.Kerry D Vandell, Walter Barnes, David Hartzell, Dennis Kraft, and William Wendt. Commercialmortgage defaults: proportional hazards estimation using individual loan histories. Real EstateEconomics, 21(4):451, 1993.Tim Westlake. Canadian cmbs still strong. Technical report, Dominion Bond Rating Service(DBRS), November 2007. URL Extracted August 2017.154Louis Wilde, Christopher Loos, and Jack Williamson. Pipelines and property values: An eclecticreview of the literature. Journal of Real Estate Literature, 20(2):245–259, 2013.Daniel T Winkler and Bruce L Gordon. The effect of the bp oil spill on volume and selling pricesof oceanfront condominiums. Land Economics, 89(4):614–631, 2013.Jeffrey E Zabel and Dennis Guignet. A hedonic analysis of the impact of lust sites on houseprices. Resource and Energy Economics, 34(4):549–564, 2012.155


Citation Scheme:


Citations by CSL (citeproc-js)

Usage Statistics



Customize your widget with the following options, then copy and paste the code below into the HTML of your page to embed this item in your website.
                            <div id="ubcOpenCollectionsWidgetDisplay">
                            <script id="ubcOpenCollectionsWidget"
                            async >
IIIF logo Our image viewer uses the IIIF 2.0 standard. To load this item in other compatible viewers, use this url:


Related Items