The Importance of Originator-Servicer Affiliation in Loan Renegotiation

Transcription

1 The Importance of Originator-Servicer Affiliation in Loan Renegotiation James N. Conklin Moussa Diop Thao Le Walter D Lima May 25, 2018 Abstract This paper presents evidence that affiliation between the mortgage servicer and the originator provides a mechanism to reduce information frictions inherent in debt renegotiation. We find that originator-servicer affiliation increases the likelihood of modification by 10%-23% using a large sample of delinquent securitized non-agency mortgages. Postmodification, affiliated loans are also 7.3% more likely to not return to severe delinquency within 12 months. Further examination reveals that affiliation affords servicers lower-cost access to borrower and loan information, thus improving their ability to implement effective debt restructuring strategies. In the absence of standardized information transmission between originators and servicers, information critical for debt renegotiation will be lost as banks disintegrate origination and servicing. JEL Classifications: G21, R2, R3 Keywords: Information Asymmetry, Mortgage Default, Debt Renegotiation, Servicing, Securitization, Mortgage Redefault, Non-agency MBS Terry School of Business, University of Georgia, jnc152@uga.edu Wisconsin School of Business, University of Wisconsin-Madison, moussa.diop@wisc.edu. Smeal College of Business, Pennsylvania State University Mendoza College of Business, University of Notre Dame, wdlima@nd.edu.

2 1 Introduction One of the most prominent concerns in the aftermath of the financial crisis of 2008 is the rise of mortgage securitization and its associated problems. In the traditional model of vertically integrated mortgage lending, the originator, the servicer and the owner of the mortgage are the same entity. In the securitization model of lending, however, these are three separate entities who may have no connection with each other, causing the three links between them to be severed: originator-owner, servicer-owner, and originator-servicer. Much of the academic research has focused on the conflicts of interest arising from the first broken link the separation between the mortgage originator and its ultimate owners (MBS investors). This disconnection results in lenders lack of incentives to carefully screen borrowers in the initial loan underwriting process (see, for example, Dell Ariccia et al. (2012), Keys et al. (2010), Kruger (2016), Mian and Sufi (2009), Purnanandam (2011)). In addition, originators can make use of their private information to cherry-pick better loans to keep in their portfolio while offloading subpar loans to MBS investors (Agarwal et al. (2012b) and Elul (2016)). Turning to the separation between the loan servicer and MBS investors, the debate on whether it gives rise to agency problems remains unsettled to date. Piskorski et al. (2010) posit that servicers may have different financial incentives to service securitized loans because they do not fully internalize the costs and benefits of their decisions. Such agency conflicts might inhibit renegotiation of troubled loans. 1 In support of this reasoning, Piskorski et al. (2010) and Agarwal et al. (2011) show that bank-held loans are significantly more likely to be renegotiated than comparable securitized mortgages. However, studies by Adelino et al. (2014) and Ghent (2011) find no differences in the renegotiation rates for securitized loans and loans held on banks balance sheets. In contrast to the abundant research mentioned above, there is little work on the potential problems associated with the third broken link that is, the detachment between originator 1 As pointed out in Eggert (2007), loan modification is labor intensive and time consuming, essentially equivalent to re-underwriting the mortgages. The cost of modification is rather expensive and the compensation structure of servicers does not cover these modification costs. When dealing with defaults, the servicer can recover foreclosure cost but not modification cost. In the recent foreclosure crisis, record default rates caused servicers to favor cost-cutting through automated foreclosure processes rather than risking incurring modification costs with a low likelihood of success. 1

3 and mortgage servicer. This paper attempts to fill this gap in the literature. Because mortgage servicers are responsible for loss mitigation efforts on delinquent loans, including debt renegotiation 2 and foreclosure, 3 servicers play a crucial role in mortgage markets during economic downturns. We posit that severing the link between the originator and the servicer a common practice in securitization inhibits the mortgage servicer s ability to effectively manage loan performance. To test our hypothesis, we investigate whether originator-servicer affiliation affects the servicer s loss mitigation decisions and their effectiveness in rehabilitating delinquent securitized loans. Specifically, we first examine whether originator-servicer affiliation affects the likelihood of a seriously delinquent loan (60+ days behind) receiving a modification. Using a large sample of non-agency securitized mortgage loans, 4 we find that loans whose servicers are affiliated with originators are 10%-23% more likely to be modified (relative to the mean) after controlling for contract and borrower characteristics, house price changes, general economic conditions, and an extensive set of fixed effects. Despite the rich set of control variables included in our regression model, there remains a potential concern that unobserved loan quality as opposed to differential ability drives the relationship between originator-servicer affiliation and debt renegotiation. In other words, servicers ability to implement loss mitigation efforts is not different for affiliated loans. Rather, the affiliated loans themselves may be different. Demiroglu and James (2012) argue that originators screen more intensively on loans where the sponsor or the servicer is likely to be an affiliated entity. Thus, affiliated loans may have higher average quality along dimensions unobservable to the econometrician, and may represent better candidates for modification. We perform three different tests to confirm that our results hold even after controlling for this potential endogeneity issue. First, we use propensity score matching to control for the 2 Debt renegotiation is typically called loan modification in mortgage markets. We will use these two terms interchangeably throughout this paper. 3 Generally, the lender has several options when dealing with delinquent loans: forebear on the delinquency, modify the loan terms, allow a short sale (whereby borrower can sell the property at a price lower than the remaining loan balance), or foreclose on the property. 4 Focusing only on securitized loans ensures that the cash flow rights of the mortgages are sold to mortgage backed security (MBS) investors regardless of whether the originator and the servicer are affiliated entities. Thus, any principal-agent issues between the MBS investors (the principal) and the servicers (the agent) should be independent of originator-servicer affiliation. 2

4 differences in observable characteristics between the two groups of loans. Second, we construct an instrument for affiliation similar to the one used by Demiroglu and James (2012) and employ 2SLS to estimate the relationship between affiliation and loan modification decisions. Finally, we exploit the changes in affiliation induced by the acquisition of several servicers in 2008 as a quasi-experiment to achieve identification. Regardless of the methods used, the positive impact of affiliation on the propensity of servicers to offer modification remains highly robust. This suggests that affiliation has a causal impact on the likelihood of debt renegotiation independent of any potential quality differences between affiliated loans and their unaffiliated counterparts. In addition, we further document that mortgages modified by affiliated servicers are less likely to redefault after modification, suggesting that affiliation also allows the servicer to renegotiate mortgages more effectively. In the last section of the paper, we propose and empirically test a potential explanation for our result. We hypothesize that heightened asymmetric information is a driving factor behind the lower modification rate of unaffiliated loans. Success in loan modification depends largely on the servicer s ability to determine the likelihood that the borrower can continue making payments after their contract is modified. Since this decision requires costly information production, the originator of a loan who has already gone through the initial underwriting process will have an advantage over other institutions performing this task. This information friction can present a major challenge to loan modification for a servicer who has no access to the originator s information. Although it is difficult to test it directly, we provide evidence consistent with this proposition. We find that 58% of affiliated servicers in our sample have information regarding the borrower s income and liabilities (debt-to-income, or DTI ratio) at the time of origination, while only 2% of the unaffiliated servicers have this information. Moreover, affiliated servicers are more likely to know about the existence of a second mortgage, thus they have a more precise combined loan-to-value ratio for the mortgage in question. 5 Taken together, the evidence suggests that affiliated servicers possess more information regarding the loan and borrower characteristics, which in turn may be useful in aiding their modification decision. 5 To be clear, we do not argue that these two pieces of information are absolutely critical to the debt renegotiation decision of the servicer. Instead, our primary purpose is to provide evidence that affiliated servicers have more information available to them than their unaffiliated counterparts. 3

5 This paper adds to the broad literature on debt renegotiation discussed in Section 2. We also add to the literature examining asymmetric information in mortgage markets and its role in the recent financial crisis. Although several papers study informational problems in mortgage securitization and origination (Mian and Sufi (2009), Keys et al. (2009), Keys et al. (2010), Purnanandam (2011), Keys et al. (2012), and Demiroglu and James (2012))), relatively little research has examined asymmetric information and its impact on debt renegotiation. 6 The paper that is most relevant to ours is Demiroglu and James (2012), who study the performance of loans whose originators and servicers are related. They conclude that the better performance of such mortgages, in terms of lower ex-post net loss and foreclosure rates, is evidence of originators selectively screening them more carefully in the initial loan origination process. In this paper, we go beyond their findings by showing that originator-servicer affiliation also causes servicers to act differently in debt renegotiation at the time of loan defaults. Our results suggest that separating the servicer and the information collector (the originator) introduces asymmetric information that prevents efficient mortgage modifications. Thus, the cost associated with mortgage securitization is not limited to the moral hazard and adverse selection problems at the beginning of the process, but extends throughout the life of the underlying mortgages. Our paper has important policy implications as well. High levels of mortgage defaults combined with low rates of mortgage modifications prompted regulatory changes in the wake of the recent mortgage crisis. Several regulatory changes, including the Secure and Fair Enforcement for Mortgage Licensing Act of 2008 (SAFE Act) and the Ability to Repay rule, target lax screening in the loan origination/underwriting process. Although these policies may increase information collection in loan screening, they do not address the problem caused by originator-servicer separation in debt renegotiation examined in this paper. Also, many of the large vertically integrated banks that traditionally handled origination and mortgage servicing have curtailed these activities in recent years, at least in part due to higher compliance and regulatory capital costs relative to non-banks (Kaul and Goodman (2016)). The results in our 6 Adelino et al. (2013) and Mayer et al. (2014) are notable exceptions. 4

6 paper suggest that this could heighten asymmetric information problems that may constrain mortgage debt restructuring in the future. 2 Related literature Due to the severity of the housing market downturn and the resulting mortgage debt crisis, mortgage debt renegotiation has received considerable attention in the academic literature. Much of this research has focused on investigating the hypothesis that securitization, which has played a central role in the massive development of the mortgage market since 2000, may represent a major barrier to debt renegotiation. 7 For the period , Agarwal et al. (2011) document that bank-held loans are 26 36% more likely to be renegotiated than comparable securitized mortgages. However, this type of study is often plagued with selection issues due the possibility that originators can choose to sell loans of lower quality to investors. To overcome this problem, Piskorski et al. (2010) make use of the early pay default (EPD) clauses 8 in many Pooling and Servicing Agreements to design a quasi-experiment around the EPD threshold, and confirm that the foreclosure rate on delinquent mortgage loans held in bank portfolios is 13% to 32% lower than on similar securitized loans originated in 2005 and However, there is not unanimity on the causal effect of securitization on loan renegotiation among researchers. In contrast to the above papers, Adelino et al. (2013) show that differences in modification rates between portfolio and securitized loans over the period were economically small and statistically insignificant. Furthermore, Adelino et al. (2014) question the validity of the EPD experiment designed by Piskorski et al. (2010) since investors seldom strictly enforce EDP clauses. Using the EPD threshold as an IV in a two-stage approach to achieve identification, Adelino et al. (2014) find no differences in the renegotiation rates for securitized loans and loans held on banks balance sheets, thereby refuting the prevalent 7 Eggert (2007) discusses several frictions preventing debt renegotiation for securitized mortgages, including agency issues related to the mortgage servicer, securitization contracts that limit servicer discretion, and conflicting interests between investors in different tranches of securitizations. 8 EPD require loan originators to repurchase any securitized loans that become delinquent typically within 90 days of being securitized. 5

7 conclusion that securitization creates frictions to loan modification. A similar conclusion is also found in Foote et al. (2010). In addition, Ghent (2011) shows that modification was rare even without securitization: less than 2% of outstanding loans received any concessionary modification during the Great Depression ( ) when mortgage securitization was almost non-existent. However, another recent attempt to settle this debate by Kruger (2016) uses a quasi-experiment to estimate that securitization increases the probability of foreclosure by 4.7 percentage points and decreases the probability of modification by 3.6 percentage points. Although it is tempting to blame securitization for the lack of modification effort by servicers, many studies have proven that there are other elements to the story (e.g., Adelino et al. (2013)). Servicer characteristics (e.g., size, specialization, see Levitin and Twomey (2011)) possibly also affect the resolution of delinquent loans. For example, Agarwal et al. (2011) show that the addition of servicer fixed effects to their model increases its explanatory power by 45% for private-label loans. Ratnadiwakara (2016) shows that servicer characteristics (capacity and access to information on borrowers) affect foreclosure rates of securitized mortgages. In summary, the debate on the effect of securitization on post-default outcomes remains unresolved in the current literature, notwithstanding several attempts to reconcile the evidence. To provide an alternative explanation for the low debt renegotiation rates, Adelino et al. (2013) develop a simple model of lenders loss mitigation decision building upon two prior papers: the cost-benefit tradeoff theory of Ambrose and Capone (1996) and the information asymmetry problem mentioned in Wang et al. (2002) that borrowers have private information about their financial conditions and willingness to repay the mortgage. In particular, modification rates will be lower when it is more difficult for lenders and servicers to evaluate borrowers ability and willingness to repay the debt. They find a negative correlation between modification rates and self-cure rates of delinquent loans, which serve as a proxy for the information problem, for the period from 2005 to Such a finding suggests that resolving this information problem should increase loan modification rates. The effectiveness of loan modifications should also be an important consideration in this debate. Recent research in this area has suggested that the probability of re-default depends 6

8 on the type and timing of modification. Among the various ways to lower monthly payments for borrowers, reducing the principal amount (which effectively reduces LTV ratio) is found to be the optimal type of loan modification, both theoretically in Das and Meadows (2013) and empirically in Quercia and Ding (2009) and Haughwout et al. (2016). 9 On the other hand, Agarwal et al. (2011) find that greater reductions in interest rates are associated with lower re-default rates. There are conflicting findings regarding the effect of extending loan duration on re-default probability, with a positive correlation found in Voicu et al. (2012) and a negative link found in Agarwal et al. (2011). Finally, Quercia and Ding (2009) assert that earlier intervention helps reduce the risk of subsequent defaults. Not surprisingly, the success rate of loan modification also depends to a great extent on the characteristics of loans and borrowers. For example, the post-modification performance of high FICO score loans, full document loans, smaller balance loans, loans with positive equity, refinance loans, prime loans, first lien loans, and fixed rate loans is superior to their counterparts (Adelino et al. (2013); Agarwal et al. (2011); Goodman et al. (2011); Haughwout et al. (2016); Quercia and Ding (2009)). Agarwal et al. (2011) also show that modifications of bank-held loans are more efficient as they have 9% lower re-default rates than mortgages underlying MBS. Interestingly, Zhang (2013) shows that the differences in self-cure and re-default rates between portfolio and securitized loans converge in the long run. Our study addresses an important missing piece in the literature: the role of originatorservicer relationship on the likelihood and success of loan modifications. Originator-servicer affiliation may proxy for loan quality (Demiroglu and James (2012)), access to information, or some unknown factor. We convincingly show that our contribution reaches far beyond the predictable effects of servicer-affiliation on mortgage modifications and re-defaults implied by Demiroglu and James (2012) and present evidence supportive of the existence of an information channel through affiliation. 9 These studies acknowledge that their small samples have limited statistical power because principal forbearance is very rare. 7

9 3 Data 3.1 Mortgage Data We use ABSNet Loan, a comprehensive non-agency mortgage origination, performance, and securitization database compiled by Lewtan, a Moody s Analytics company. Lewtan sources, normalizes, and analyzes vast amounts of non-agency mortgage data reported by MBS trustees and servicers to provide granular information on a broad array of mortgage loan and deal attributes under the ABSNet Loan brand. This database contains more than 22 million loans collateralizing roughly 7,000 MBS deals. 10 Using monthly loan performance updates, we identify loans that were seriously delinquent in December We then track first-time modifications over the next 12 months, January through December Following an approach similar to Maturana (2014), we focus our analysis on a short window to avoid contamination from various government programs aimed at curbing foreclosures during the recession following the housing market bust. Expanding the study period complicates matters because modifications may have come about as a result of federal programs, such as the Home Affordable Modification Program (HAMP) enacted in March 2009 to provide relief to distressed homeowners. Moreover, it is not clear if or how these programs would differentially impact affiliated versus unaffiliated servicers. We chose the end of 2007 as the onset of the likelihood of modification for all of our loans because this date corresponds to a time when modifications were being discussed intensely as a potential solution to high rates of foreclosure (FDIC (2007)). Also, using remittance reports from 2007 and 2008 for twenty-six pools of subprime mortgages originated in 2005 and 2006, White (2009) shows modification rates were relatively modest up until December of For these reasons, 10 Lewtan markets its services primarily to non-agency MBS investors for valuation purposes. The bulk of the loans in the ABSNet database were originated after In addition to ABSNet Loan, the company provides periodic collateral valuations it markets as ABSNet HomeVal. 11 Lewtan monitors changes in loan terms monthly in a separate data file within ABSNet Loan. The company receives monthly loan modification updates directly from servicers and trustees, but it also independently checks modifications against changes in loan characteristics recorded in the monthly loan snapshots for confirmation. Unreported modifications identified by the company are then recorded and flagged as implied modifications. We classify multiple reports of modifications within the same month as one modification. 8

10 in this paper we focus on modifications in 2008, but our findings remain unchanged when we expand the study period through Our initial sample consists of 632,498 loans collateralizing MBS deals issued from 2004 to 2007 that have current servicer information at the end of 2007 and that are at least 60 days delinquent at that date. We apply the following filtering criteria to the original sample. First, we drop second liens, HELOCs, loans where the subject property state is missing, loans originated in Puerto Rico and Guam, and loans originated before Next, we exclude loans with missing originator information. After further cleaning of the data to deal with potential reporting errors, our final sample consists of 422,234 loans collateralizing 1,554 nonagency MBS deals. 13 These loans were originated by 985 institutions and managed by 56 servicers at the end of Table 1 lists the 20 largest originators and servicers, which together account for 84% and 94% of our sample, respectively. The mortgage servicing industry experienced considerable consolidation in Table 2 shows the fates of the 20 largest servicers in our sample. 14 We will discuss this further in Section where we use servicer acquisitions as a quasi-experiment to provide identification for the servicer affiliation effect. 3.2 Servicer Affiliation and Descriptive Statistics Our primary independent variable of interest is a dummy that equals one if the servicer is affiliated to the initial information collector (the originator). To construct this variable we match the originator of each loan to the current servicer in December 2007 and then manually check whether the two entities are related. We recognize that originator-servicer affiliation is dynamic, and the further our study period extends the noisier our measure becomes. We believe this is not a major concern for our study for two reasons. First, if our measure is noisy, it should bias us away from finding a significant relationship between affiliation and 12 We excluded loans originated before 2000 in order to focus on the last housing boom. 13 To limit the potential for reporting errors, we dropped loans with a balance less than $25,000, loans with original property appraised values missing or less than $25,000, loans with original LTV below 25% or above 150%, loans with less than 60 months of remaining life at the end of 2007, loans with initial maturity of more than 600 months, missing borrower locations (CBSA), and observations with missing loan characteristics included in equation (1). 14 In a bank failure, the mortgage servicing assets are typically transferred to another servicing company. This is analogous to a servicing firm being acquired. 9

11 modification. Second, we limit our study period to a relatively short period of time after we measure affiliation (6 months and 12 months), thus reducing the possibility that our proxy for affiliation is stale. Table 3 shows that loans from affiliated originators represent 63.3% of our sample. This reflects the fact that large financial institutions generally maintain an integrated mortgage origination and securitization business in order to capture the entire value chain. There are economically and statistically significant differences between the two groups in servicers unconditional propensity to modify delinquent loans over 12 months, with unaffiliated loans 0.7% less likely to be modified. The redefault rate of modified loans, measured over the 90-day postmodification period, is slightly higher for affiliated loans, as shown in the second row of Table 3. In Section 4, however, we find that our conditional estimation yields the opposite result. In terms of borrower and loan characteristics, even though the differences between the two groups of loans appear statistically significant, they are mostly economically negligible. Most importantly, we note that there is no conclusive evidence pointing to any discernible differences in their risk profile. Although affiliated loans have higher FICO scores, lower interest rates, and lower DTI ratios on average, which suggest higher quality borrowers, affiliated loans are also more likely to have negative amortization and low or no income documentation, and have higher CLTVs, all of which are indicative of high risks. Thus, it is unclear from these observable characteristics that the two groups of loans differ in their quality. 4 Empirical Results 4.1 Modification Decision To investigate if servicers loss mitigation decisions depend on whether a loan was originated by an affiliated entity, we estimate a model of the following form: Pr(Modified ik ) = α + β Affiliation k + X iγ + Z iδ + Year i + Location l + Servicer k + ξ i, (1) 10

12 The dependent variable Pr(Modified ik ) is the probability that mortgage i, serviced by entity k, is modified in 2008, conditional upon being 60 or more days delinquent in December Affiliation k is an indicator that equals one if the servicer of the loan is affiliated with the mortgage originator. X i is a vector of loan characteristics that includes the delinquency status at the end of 2007, loan seasoning (number of months between loan origination and securitization), income documentation indicators, an Alt-A loan indicator, amortization type indicators (negative amortization, balloon, and interest-only), an interest rate type indicator (ARM), a purchase (versus refinance) indicator, a single-family house indicator, an owner-occupied unit indicator, the loan amount, the interest rate, the amortization term, the borrower s credit score, the initial combined loan-to-value ratio (CLTV), and an estimate of the current loan to value ratio (LTV). 15 As Table 2 suggests, many servicers became distressed in 2008 as mortgage delinquencies rose to unprecedented levels. Since modifications are labor intensive, it is possible that distressed servicers became capacity constrained in 2008, leading to lower rates of mortgage modifications. On the other hand, since servicers are still required to advance payments to MBS investors while a borrower is delinquent, and are not reimbursed for this expense until after a delinquency is resolved (e.g., foreclosure, modification, mortgage brought current), distressed servicers may be more likely to perform mortgage modifications. Therefore, X i also includes a proxy for servicer distress: the fraction of the servicer s portfolio that is delinquent in the previous month. 16 Z i is a vector of controls for housing market and general economic conditions, including changes in house prices, monthly inflation, mortgage interest rates, area income, and area unemployment. 17 We also include a set of loan origination year fixed effects (Year i ) location fixed effects, and servicer fixed effects (Servicer k ). 18 For convenience, we will primarily assume a linear probability model (LPM) and use OLS to estimate the model s parameters given the number of fixed effects. However, this choice of estimation 15 Current LTV= balance t value t, where balance t is the current outstanding balance of the mortgage and value t is the updated value of the house after adjusting for changes in the FHFA MSA-level house price index. 16 We thank an anonymous referee for suggesting this proxy for servicer distress. 17 Time-varying covariates in Z i are measured at the last date before an event. In equation (1) the event is modification, so they are measured as of the month prior to modification. For loans that are not modified, they are calculated at the end of our sample period. In Section 4.3 redefault is the event. 18 Agarwal et al. (2012a) show that some servicers have the organizational structure that is capable to handle large-volume renegotiation cases, while others do not have this capability. We therefore add servicer fixed effects to control for servicer-specific factors that may affect renegotiation rates. 11

13 method does not drive our empirical findings. β is our primary estimate of interest: β > 0 suggests that affiliation increases the probability of debt renegotiation. Column 1 in Table 4 reports the results from OLS estimation of the probability of modification over 12 months for our sample. Consistent with our main hypothesis, originator-servicer affiliation is significantly positively related to the likelihood of modification. The economic magnitude of the relationship is large as well. The model specification in column 1 shows that affiliation increases the probability of modification by 2.8% in absolute terms with the control variables behaving as expected. Relative to the average rate of modification, this represents a 23% increase in the likelihood of modification. 19 Column 2 includes our proxy for servicer distress and shows that distressed servicers are significantly less likely to modify mortgages. Affiliation remains positively related to the likelihood of mortgage modification in column 2. Evidence suggests that significant heterogeneity exists in loss mitigation strategies across servicers (Stegman et al. (2007), Agarwal et al. (2011), Agarwal et al. (2012a), and Collins et al. (2016)). This heterogeneity in modification strategies could be due to technology (i.e., cost structure), size, expertise, or simply business strategy. Thus, we include servicer fixed effects in our estimation reported in column 3 to exploit within servicer variation in affiliation. In this specification, loans from affiliated originators are on average 1.5% more likely to be modified over 12 months, which represents a 12% increase relative to the mean. 20 As the mortgage crisis unfolded, it became increasingly difficult for mortgage servicers to put-back recently originated non-performing loans to financially distressed (or failed) originators. The inability to put-back loans could affect the servicer s loss mitigation strategies. Thus, in column 4 we include servicer-origination year fixed effects to account for this possibility. 21 Affiliation remains positively related to the likelihood of modification in column 4. Since the economic magnitude of the coefficient on affiliation is smallest (9.7% relative 19 This is calculated as the marginal effect divided by the 12-month mean modification rate in our sample 2.8%/12.1%. 20 Adding originator fixed effects does not materially change our findings. To the contrary, it pushes the magnitude of the affiliation effect back to that of column 1. The effect of affiliation on the likelihood of loan modification also remains statistically significant when we cluster standard errors by year of loan origination, originator, or location (CBSA). 21 An alternative estimation strategy to account for this possibility is to include loan originator-origination year fixed effects. Our results are not materially affected when we use this alternative specification. We thank an anonymous referee for this suggestion. 12

14 to the mean modification rate for the treated of 12.1%) when we include servicer-origination year fixed effects (column 4), we adopt this as our main specification in the remainder of the paper to present the most conservative estimates of the effect of affiliation. In summary, the results in Table 4 provide consistent evidence that originator-servicer affiliation strongly affects the probability of debt renegotiation. Depending on the specification, servicers are 10% - 23% more likely to modify affiliated loans, even after we account for an extensive battery of control variables and a rich set of fixed effects. 4.2 Endogeneity In the previous section, we implicitly assumed that originator-servicer affiliation is exogenously determined. However, this may not be an accurate depiction of reality. Originators may take advantage of their private information to cherry-pick low risk loans to retain in their servicing portfolios. It follows that affiliated loans may be systematically different from unaffiliated loans, and affiliation may therefore be endogenous. In support of this reasoning, Demiroglu and James (2012) compare ex-post foreclosure rates and find a lower rate among loans whose originators are affiliated with their servicers, suggesting that they carry lower risks. 22 If this is true, our finding that affiliated loans are modified more often may simply be attributed to their ex-ante superior quality. In this section we perform three tests to confirm that the positive effect of originator-servicer affiliation on loan modification remains robust even after controlling for this endogeneity issue. First, we use propensity-score matching to mitigate the differences between affiliated and unaffiliated loans. Second, we identify an instrumental variable and estimate our model using two-stage least squares. Third and most importantly, we make use of the acquisition of several servicers in 2008 as a quasi-experiment to infer the causal effect of affiliation on modification decisions. 22 Although the summary statistics reported in Table 3 do not suggest any economically discernible differences between the two groups of loans, we cannot rule out that they differ in characteristics unobservable to econometricians. 13

15 4.2.1 Propensity Score Matching Although we control for borrower characteristics in the previous estimations, the estimated effect of affiliation on modification might partly reflect differences in characteristics between affiliated and unaffiliated loans if the models are not correctly specified, as discussed in Demiroglu and James (2012). A common approach to address this issue is propensity-score matching (PSM) analysis, which directly mitigates the confounding effects of differences in observable characteristics of affiliated and unaffiliated loans (Rosenbaum and Rubin (1983)). The propensity-score matching technique involves matching affiliated and unaffiliated loans based on observable characteristics and then examining the effect of affiliation on modification in the matched sample. We compute propensity scores for our original sample as the loans likelihood of affiliation using a probit regression relating affiliation to loan characteristics. We then use the loans propensity scores (estimated likelihoods of affiliation) to find a matching unaffiliated loan for each affiliated loan using a nearest neighbor matching technique. We conduct the matching with replacement and impose a common support restriction when estimating the treatment effect of affiliation on modification. Panel A of Table 5 reports the estimated coefficients of the explanatory variables from the probit model of the likelihood of treatment (affiliation). In propensity score matching only pretreatment variables are included in generating the propensity for treatment. Thus, we include loan characteristics at origination and location fixed effects to control for heterogeneity in origination practices in our probit model. Panel A shows a clear relationship between the likelihood of affiliation and loan characteristics. We note that the regression results do not necessarily match with the differences in the unconditional means reported in the summary statistics in Table 3. Interestingly, the estimated coefficients provide mixed evidence on our prior expectation that affiliated loans have lower risks. For example, while the signs of the coefficients on loan amount, CLTV, and credit score are inconsistent with the predictions of this hypothesis, other variables such as interest rate, negative amortization indicator, and balloon indicator have the correct signs. 14

16 Our matched sample is created using a nearest neighbor matching technique based on the propensity score from the first stage along with exact matching on loan originator. Thus, each treated (affiliated) observation has a similar control (unaffiliated) observation originated by the same lender. 23 In Panel B, we report the treatment effect of affiliation on modification on the matched sample using the model in Column (4) of Table 4. The estimated average treatment effect is slightly smaller than the effect estimated from the unmatched sample (0.009 versus ), but is still economically and statistically significant. Affiliation increases the likelihood of modification by 7.4% relative to the mean modification rate for the treated (12.1%). Overall, the outcome of the propensity-score matching analysis shows that the documented effect of affiliation on modification is not due to differences in loan characteristics between affiliated and unaffiliated loans Instrumental Variable In our second approach to address the endogeneity problem, we employ the two-stage least squares (2SLS) method which is commonly used in the literature to estimate causal effects. To use this approach, we need to find an instrument that meets two requirements. First, the instrument must be highly correlated with affiliation, conditional on other covariates. Second, the instrument must meet the exclusion restriction that it only affects modification through its impact on the likelihood of affiliation. Adopting the approach used by Demiroglu and James (2012), we construct an instrument that measures for each loan in our sample the originator s share of mortgages serviced by affiliated entities over the three months prior to that mortgage s origination month. We argue that this instrument likely satisfies the two requirements for the following reasons. If a large share of the originator s loans over the past few months were sent to an affiliated servicer, there is a high probability that the next loan originated will also go to an affiliated servicer. However, the share of the originator s business that went 23 We implement the matching using psmatch2 in Stata (Leuven and Sianesi, 2018). The matching considerably reduces differences in characteristics between affiliated and unaffiliated loans in our PSM sample compared to the unmatched sample. However, following a common practice in the literature, we also include the matching variables in our treatment estimation model to control for remaining minor differences in characteristics between affiliated and unaffiliated loans. 15

17 to an affiliated entity before origination should not directly impact the servicer s subsequent modification decision on an individual loan. Table 6 reports the coefficients from the first and second stage of the 2SLS model using two different specifications. The null hypothesis that affiliation is exogenous is rejected at the 1% level of confidence using the score test from Wooldridge (1995). From the first stage results, the instrument for affiliation in both specifications is significantly positively related to affiliation and we are able to reject the null hypothesis that our instrumental variable is weak. Turning to the second stage results on the relationship between originator-servicer affiliation and modification, we find that affiliation remains significantly positively related to the probability of modification. The marginal effects of affiliation from the 2SLS estimations presented in Panel B of Table 6 is identical to that from the OLS estimations in Table 4, confirming again that originator-servicer affiliation has a causal effect on the likelihood of modification, even after controlling for the potentially endogenous nature of affiliation Quasi-experiment Our final and most important strategy to identify the causal impact of originator-servicer affiliation on loan modification exploits a quasi-experiment created by the consolidation of the servicing industry during the financial crisis. In particular, several servicers were acquired or went out of business during our sample period (see Table 2). After acquisition (or failure), any link between the originator and the servicer is severed. Thus, acquisitions represent exogenous shocks to the affiliation status of the loans in our sample. We exploit servicer acquisitions and failures within a difference-in-differences framework to estimate the causal impact of affiliation on modification. Demiroglu and James (2012) use a similar methodology to identify the causal impact of originator-sponsor affiliation on the likelihood of mortgage default. We focus our analysis on loans serviced (at the end of 2007) by firms that were acquired or failed during The list of acquired (failed) servicers consists of the following firms: Countrywide, EMC, WAMU, Equity One, Option One, Indymac, Fremont Investment & Loan, and American Home Mortgage Servicing. Loans serviced by Avelo Mortgage are not included 16

18 in this sample since Goldman Sachs owned the servicing rights before and after the merger with Litton Loan Servicing. Also, we exclude loans serviced by Home Loan Services, Inc. since the firm was not acquired until the last day of our sample period. In our context, the intuition of the difference-in-differences strategy is to determine whether the difference between affiliated and unaffiliated loans changes after affiliation is severed (post-acquisition). To examine the difference-in-differences, we estimate a model of the following form: Pr(Modified) = α + β 1 Affiliation + β 2 Post-Acquisition + β 3 Affiliation x Post-Acquistion + X γ + Z δ + Location + Servicer-Origination Year + ξ, (2) where Post-Acquisition is a dummy variable equal to one if the modification date (or the last date the loan is observed for non-modified loans) occurs after the servicer is acquired. The interaction term (Affiliation x Post-Acquistion) indicates whether affiliation is exogenously severed by an acquisition (or failure). All other variables are as defined in equation (1) with subscripts suppressed for notational convenience. We again stress that the estimation sample only includes loans serviced (at the end of 2007) by firms that were acquired or failed during β 1 and β 3 are our estimates of interest in equation (2). β 1 will contain two effects: a) affiliation between the originator and the current servicer and b) other differences between affiliated and non-affiliated loans (e.g., quality). β 2 will enable us to separate the effects contained in β 1. Since the interaction term captures the termination of affiliation, β 3 will give an estimate of the causal impact of affiliation on the likelihood of modification. β 1 + β 3 will measure the differences in modification rates on affiliated loans due to other factors (e.g., quality). Table 7 presents coefficient estimates from the linear probability model of equation (2). Consistent with our previous findings, affiliated loans are significantly more likely to be modified relative to unaffiliated loans. However, termination of affiliation reduces the likelihood of modification by 2.5 percentage points, which suggests that 63% of the impact of affiliation 17

19 on modification is caused by affiliation between the originator and the current servicer. 24 The remainder of the affiliation effect (β 1 + β 3 = 0.015) is attributable to other factors such as unobservable differences in loan quality. There are two key takeaways from Table 7. First, loans that are originally affiliated are more likely to be modified even after originator-servicer affiliation is broken. This is consistent with Demiroglu and James s (2012) argument that originator-servicer affiliation proxies for differences in loan quality that are unobservable to the econometrician. The second and more important takeaway from Table 7 is that affiliation between the originator and the current servicer increases the likelihood of modification. In summary, this section presents three robustness tests to confirm that the positive effect of affiliation remains significant and consistent even after we control for the potential endogeneity issue. We can therefore conclude that affiliation causes servicers to treat loans differently when making modification decisions, for reasons independent of any quality differences between affiliated and unaffiliated loans. This finding represents a key contribution of our paper to the existing literature. 4.3 Affiliation and Loan Redefault It is important to recognize that the objective of loan modification is minimizing loss by making troubled loans perform again. Hence, we next ask if originator-servicer affiliation impacts the effectiveness of the loan modification. Specifically, we examine if the redefault rate of affiliated loans is significantly different from their unaffiliated counterparts. We estimate the following linear probability model of redefault for loans that received a modification: Pr(Redefault) = α + β Affiliation + X γ + Z δ + Year + CBSA + Servicer + ξ, (3) The dependent variable equals one if the borrower becomes 90 or more days delinquent on their mortgage within 12 months following modification. The model controls for all variables 24 This is calculated as 0.025/0.040 = 63%. 18

20 listed in equation (1) with subscripts suppressed for notational convenience. 25 The estimates in Table 8 show that after modification, affiliated loans are 7.3% more likely than unaffiliated loans to not fall into severe delinquency status again over the next 12 months. 26 Although originator-servicer affiliation is positively associated with post-modification loan performance, we note that the majority (73.2%) of modified loans redefault regardless of servicer affiliation. Therefore, we make no claims regarding the overall welfare effects of modification to borrowers and investors, only that servicers modify more effectively when the loan is from an affiliated originator. Coupled with the result from the previous section, these findings lead us to conclude that affiliation has a positive impact on not only modification volume but also effectiveness. 5 Affiliation and Information Asymmetry 5.1 Information Hypothesis of Mortgage Servicing Recent evidence suggests that information asymmetry inhibits debt renegotiation. Adelino et al. (2013) argue that borrowers have an informational advantage over servicers regarding their prospects of repayment, which can lead to an inefficiently high level of renegotiations if servicers modify delinquent loans that would have self-cured on their own. On the other hand, servicers may inefficiently delay foreclosure by modifying loans that will redefault. Furthermore, if it is costly or difficult to determine who truly needs a modification, a liberal modification policy may induce borrowers to strategically default in order to receive a mortgage modification. 27 Mayer et al. (2014) provide empirical support that this moral hazard problem is real and economically meaningful. 28 Taken together, these results suggest that information problems play a key role in the servicing of delinquent mortgages. 25 Since the redefault decision is under the purview of the borrower, rather than the mortgage servicer, we exclude the servicer distress proxy in this regression. However, the results are materially unchanged when we include the servicer distress proxy. 26 This is 1.95% divided by 1 minus the average 12-month redefault rate of 73.2%. 27 Bourgeon and Dionne (2013) theoretically examine some of these same issues to explain bank refusal to renegotiate debt. 28 In fact, in a recent contribution Goodstein et al. (2017) find evidence for a contagion effect in strategic mortgage defaults. 19

21 It follows that the availability of additional information to servicers can reduce these problems and affect their debt renegotiation strategies. However, the information available to servicers on the loans they service is not standardized. The loan boarding process illustrates this point. During loan underwriting, the originator collects information relevant to the credit and pricing decisions. Although some of the information may be standardized, such as those captured on the Uniform Residential Loan Application, additional information collected varies across originators. After origination, a subset of the information is transferred from the originator s system to the servicer s system, a process known as loan boarding. For example, the information boarded typically includes borrower contact information, loan information (e.g., loan amount, maturity date, interest rate, loan term, loan type, adjustable-rate reset parameters) and lien information (Spoto (2014)). Additional borrower and loan information required or available for loan boarding varies across servicers and originators. For example, one mortgage servicer lists 14 key fields used in mortgage boarding, but also states [W]e ll be happy to take as many as 148 fields of primary loan data. 29 Discussions with industry participants further confirm that the amount of information transferred between the originator and the servicer varies significantly. Thus, information available to servicers to aid in mortgage loss-mitigation efforts also varies both across and within servicers. 30 We hypothesize that originator-servicer affiliation reduces the information asymmetry inherent in debt renegotiation. In order to determine the most appropriate loss-mitigation strategy, the servicer must engage in costly information production. However, if some or all of this information was originally produced by an affiliated originator, obtaining this information from the affiliated entity rather than engaging in new information production reduces costs. Even if a servicer can obtain the same information from an unaffiliated originator, costs of information transmission across unaffiliated entities is likely higher than within affiliated entities. Hence, our information hypothesis posits that a rational servicer would focus its loss-mitigation efforts on loans originated by affiliated entities. Moreover, additional information will allow the ser- 29 Available at 30 Within servicer variation in boarded information is likely if the servicer works with multiple loan originators. Technically, we do not fully observe the information available to servicers; we have access only to information reported by the servicers to the MBS trustees. 20