Identifying the Effect of Securitization on Foreclosure and Modification Rates Using Early-payment Defaults ∗ Kristopher Gerardi‡ FRB Atlanta

Manuel Adelino† Duke’s Fuqua School of Business

Paul Willen§ FRB Boston

April 1, 2013 Abstract This paper develops and estimates an instrumental variables strategy for identifying the causal effect of securitization on the incidence of mortgage modification and foreclosure based on the early-payment default analysis performed by Piskorsi, Seru, and Vig (2010). Estimation results show that securitized mortgages are more likely to be modified and less likely to be foreclosed on by servicers. These results are consistent with the interpretation in Adelino et al. (2013) that low modification rates are not the result of contract frictions inherent in the mortgage securitization process.

∗

We thank Chris Cunningham, Chris Foote and Sam Kruger for thoughtful comments and Neil Desai for excellent research assistance. † [email protected] ‡ [email protected] § [email protected]

1

1. Introduction According to a popular narrative of the U.S. foreclosure crisis, by blocking renegotiation or, as it is known in the industry, loan modification, securitization has caused millions of unnecessary, socially wasteful foreclosures. Is this narrative right? Proponents point to institutional evidence, but such evidence is, by its nature, inconclusive. For example, some securitization agreements limit the number of loans that a lender can modify, but skeptics point out that these limits are rarely ever binding. Seeing the limits of the institutional evidence, researchers have turned to the data and have argued that the fact that lenders are less likely to modify and more likely to foreclose on securitized mortgages than on loans held on banks balance sheets (“portfolio” loans) proves that securitization causes inefficient foreclosures.1 There are, however, two problems with this argument. The first is that, as Adelino, Gerardi and Willen (2013) (hereafter AGW) point out, in dealing with troubled loans, the goal of the lender is not to minimize the number of foreclosures or to maximize the number of modifications, but rather to minimize losses. AGW show that a policy that leads to fewer modifications and more foreclosures can lead to smaller losses for investors, which highlights the fact that minimizing foreclosures and minimizing losses are two very different objectives. But even if one measures the success of loss mitigation by counting how many modifications or how few foreclosures the lender performs, there is another serious problem with interpreting differences between the performance of securitized and portfolio loans as a treatment effect of securitization. The problem stems from the fact that the decision to securitize a loan is not random. In this paper, we consider the problem of the selection and treatment effects of securitization, exploring a potential solution first proposed by Piskorski, Seru and Vig (2010) (hereafter PSV). The specific selection issue here is that lenders choose which loans to place into securities, and this can lead to adverse selection problems either because the lender chooses loans of unobservably worse quality or because the lender puts less unobservable effort into underwriting loans it knows it will sell.2 Suppose, for example, that lenders have private information about whether borrowers are speculators and place the speculative loans into securities but, for loss mitigation purposes, treat the loans identically.3 When house prices fall, speculators are more likely to walk away from the loans than the typical 1

Adelino, Gerardi and Willen (2013) argue that until heavy government involvement in mortgage servicing in the fall of 2008, securitization had little or no effect on the likelihood of a modification. 2 Although prior research has portrayed this as a problem of moral hazard (for example Keys et al., 2009, Mian and Sufi, 2009), strictly speaking, there is no moral hazard problem in underwriting loans. Since the lender has private information either about the quality of the loan or the effort expended in underwriting prior to contracting, adverse selection is the only potential problem. 3 See Elul (2009), Krainer and Laderman (2009) for evidence that there are unobservable differences in the performance of securitized and portfolio loans.

2

borrower, avoiding any contact with the lender and, as a result, are less likely to obtain modifications and more likely to lose their properties to foreclosure. Analysis of the data would show that lenders are less likely to modify and more likely to foreclose on securitized loans, but concluding that securitization prevented renegotiation in this scenario would be wrong. To address the selection problem, PSV propose to exploit early payment default (EPD) clauses that call on issuers to repurchase securitized mortgages if the borrower defaults in the first 3 months after securitization. PSV argue that whether the borrower first misses a payment in month 3 or month 4 after securitization is close to random and, therefore, loans that become delinquent for the first time in month 3 are randomly more likely to be on a bank’s balance sheet (because they were repurchased) than loans that default for the first time in month 4. Thus, if one restricts attention to loans that default in either month 3 or month 4, one has a plausible instrument for securitization. Figure 1 shows that although there is no discontinuity, there is a decrease in the the probability of repurchase between loans that default in month 3 versus month 4 and we show, using standard tests, that EPD is, in fact, strongly correlated with the subsequent securitization status of a loan. In this paper we use EPD as an instrument for the securitization status of a loan, and find that securitization leads to more modifications and fewer foreclosures. Figure 2 shows the probabilities of foreclosure and modification respectively by the timing of the first missed mortgage payment. One can interpret these figures as the reduced form of the instrumental variables (IV) regressions of foreclosure and modification on securitization status, and they show pronounced reductions in foreclosures and increases in modifications for the loans most likely to be securitized. Our preferred specification, a bivariate probit, implies that securitized loans are approximately 17 percent less likely to be foreclosed upon and approximately 5 percent more likely to be modified than portfolio loans. Our results differ sharply from PSV’s results, who utilize EPD and report that securitization leads to more foreclosures relative to portfolio loans. The difference in findings between the papers results from the fact that PSV do not use EPD as an instrument for securitization, but instead use it to construct what the authors refer to as a “quasi-experiment.” Rather than comparing all loans that defaulted in month 3 versus month 4 in a standard two-stage implementation, PSV compare loans that defaulted in month 3 and were repurchased by the lender with loans that defaulted in month 4 and remained securitized. By adopting this alternative strategy, PSV introduce a different selection problem into the estimation. The new selection problem occurs because investors do not force issuers to repurchase all or even most loans that default before the 3-month EPD threshold. In fact, Figure 1 shows that lenders select for repurchase only about 4 percent of the loans that become delinquent 3

before the fourth month following securitization. In short, PSV’s use of EPD to identify the effect of securitization does not address the selection problem any more convincingly than regressions that use securitization as the right hand side variable of interest. Figure 3 shows the different approaches to identification and compares them with Angrist’s (1990) classic use of IV to identify the effect of military service on future earnings. Angrist used the draft lottery as an instrument for military service, as men receiving low draft numbers were much more likely to serve in the military than men receiving high draft numbers. In the top panel of the figure we consider running simple regressions where one uses a dummy for securitization and military service, respectively. However, both regressions face the problem that neither securitization nor military service are randomly assigned, and thus are prone to selection bias. The second panel shows the IV solution in both cases, which is to find a variable that is randomly assigned, but that is also correlated with the variable of interest. In the Angrist case it is the draft lottery number, and in our case it is the time of first delinquency relative to the 90-day EPD threshold. The third panel illustrates why the PSV approach is not IV and does not solve the endogeneity problem. Effectively, what PSV do is as if Angrist had simply compared the earnings of workers with low draft numbers who served in the military and workers with high draft numbers who did not serve. Of course the problem with such an approach is that it only works if individuals were randomly assigned to military service, and if that were the case, there would be no need for the instrument in the first place. How should one interpret our results? At face value, the results show that not only did securitization not block renegotiation, it actually spurred it on. While many commentators have stressed the institutional reasons for why securitization presents problems for modifications, there are actually many reasons to think the opposite. AGW present an array of institutional evidence along these lines, including the fact that lender’s own filings with the SEC show that foreclosures reduce the value of their servicing rights whereas modifications increase them. With that said, we are skeptical of the value of EPD as an instrument for securitization. There are two main problems with the approach. First, loans that default very early in their life are different from the rest of delinquent mortgages, so it is unclear that the results using this subsample of loans can be generalized to the whole population. Second, many lenders repurchase EPD loans with the express intention of placing them in other securitization pools; of the EPDs in our sample, our data shows that more than 60 percent ended up in new securities. In our view, the correct interpretation is that this strategy is really comparing loans randomly assigned to two different flavors of securitization, rather than securitized versus portfolio loans. The paper proceeds as follows: In section 2 we review the literature on the topic of 4

securitization and mortgage outcomes, paying special attention to three papers that directly address the question of whether securitization impedes efficient mortgage renegotiation. In section 3 we present and discuss the merits of the PSV “quasi experiment,” and in section 4 we develop and estimate an alternative IV framework that uses EPDs to identify the effect of securitization on foreclosure. Finally we offer concluding remarks in section 5.

2. Existing Literature Using the same nationally representative mortgage dataset, which comes from the data provider Lender Processing Services (LPS), PSV and AGW both looked at the performance of seriously delinquent mortgages, defined as loans that were at least 60 days past due, in the period 2005 through 2008. The difference between the two papers is largely methodological. PSV compared foreclosure rates for privately securitized mortgages and portfolio loans, and interpreted the difference in foreclosure rates that they found as stemming from differences in the ability of the servicers of both loan types to renegotiate with borrowers. As we mentioned above, PSV found that the foreclosure probability for a privately securitized mortgage was between 3.8 and 7 percentage points higher than a portfolio loan, depending on the sample. Their empirical strategy was basically two-fold. In the first part of the analysis they estimated logit regressions of the probability of foreclosure on a variable that indicated whether or not the loan had been securitized and a set of control variables that included many of the key underwriting variables like credit score, loan-to-value ratio (LTV), maturity, etc. Recognizing that securitization is an endogenous variable in that it reflects a choice by the mortgage lender to either sell the loan to a securitizer or to retain the loan on its own balance sheet, in the second part of their analysis, PSV develop the “quasi experiment” that we discussed above, and which we will discuss in much more detail below. The results of the “quasi experiment” are largely consistent with the results from their simple logit regressions. Whereas PSV compare foreclosure rates for securitized and portfolio loans, AGW focus directly on modification rates. They use an algorithm to identify cases in which loans were modified in the LPS dataset, which basically uses information on whether certain loan terms change over the life of a mortgage in a way that is not stipulated by the original terms of the contract.4 The authors estimate logit regressions of the probability that a loan is modified on a variable that indicates whether or not the loan had been securitized and a 4

The LPS dataset does not contain direct information on modifications. That is the servicers that provide the data on loan performance do not disclose whether they have given a mortgage borrower a modification.

5

set of control variables similar to the set used by PSV. AGW found that both portfolio and securitized mortgages are characterized by very low modification rates and found no evidence that portfolio loans are more likely to be modified than securitized loans. The authors also recognize that the endogeneity of the securitization decision could bias the estimation results from the logit regressions. To try to address this issue, they re-estimate the logits on subsamples of mortgages that they argue are less likely to be characterized by unobserved heterogeneity (for example, loans that are not eligible to be purchased by the GSEs and loans for which borrowers fully document their income). In trying to reconcile their results with the PSV results, AGW argue that differences in foreclosure rates likely reflect more than just differences in the propensity to renegotiate with delinquent mortgage borrowers. For example, higher foreclosure rates for portfolio loans could reflect the reluctance of portfolio lenders to recognize losses for accounting purposes. As a result, interpreting foreclosure rate differences to be reflective of differences in the likelihood of loan renegotiation may be misleading.5 Another recent paper by Agarwal, Amromin, Ben-David, Chomsisengphet and Evanoff (2011) attempts to reconcile the differences between the AGW and PSV analyses. The Agarwal et al. (2011) paper also considers differences in renegotiation rates across privatelabel and portfolio mortgages, but uses a different data source and a different sample period. These authors find statistically significant differences in renegotiation rates between securitized and portfolio loans, with portfolio loans being less likely to be modified, but the economic magnitude of the effect is small and the sample period is contaminated by significant government intervention in the mortgage market. We defer to AGW for a more detailed discussion of the Agarwal et al (2011) paper.6 5

A nice illustration of this issue can be seen in the comparison of judicial and power-of-sale foreclosure states. At any given point in time, there are large differences in foreclosure rates between the two types of states (Mian, Trebbi, and Sufi (2011) and Gerardi, Lambie-Hanson, and Willen (2011)). However, Gerardi et al. find no differences in modification rates between the two states, and find that the gaps in foreclosure rates between the two types of states close over time, so that cumulative foreclosure rates over a period of several years are roughly comparable. If the PSV interpretation of differences in foreclosure rates is applied to this context, one would wrongly conclude that there are significant differences in the likelihood of renegotiation between the two types of states. 6 It is important to note that mortgage servicers and lenders have other options besides foreclosure and modification when it comes to dealing with delinquent mortgage borrowers. For example, a lender could agree to a short-sale or a deed-in-lieu of foreclosure with a delinquent borrower, which would force the borrower to move out of the property, but would avoid the lengthy delays that typically characterize foreclosure proceedings. While there is anecdotal evidence that short-sales have become more prevalent since the beginning of the foreclosure crisis, there is little empirical evidence on these the prevalence of these loss mitigation alternatives, and no evidence, to our knowledge, on whether there are important differences in the frequency of these alternatives across securitized and portfolio mortgages.

6

3. Early-Payment Default as an Instrument for Mortgage Securitization The study of the impact of securitization on mortgage outcomes inevitably runs into the problem of omitted variables bias — even when observables are taken into account, it is possible that the performance of mortgages is affected by unobserved factors that are correlated with both mortgage outcomes and the probability that a loan is securitized. To deal with this issue, PSV develop a strategy which involves using early-payment defaults as a “quasi-experiment” for whether a loan remains in a securitization trust (and is serviced as a securitized loan) or gets transferred to the mortgage servicer’s balance sheet and is subsequently serviced as a portfolio loan.7 The empirical strategy employed by PSV uses the exact timing of delinquency to assign each loan to the treatment or the control group. The authors argue that securitization trusts that include an EPD clause specify that loans becoming delinquent in the first 90 days after securitization are subject to repurchase, while those that become delinquent after 90 days are not.8 Based on this, PSV create a “treatment” group that includes loans that become 30 days delinquent in the third month after securitization, transition to 60 days delinquency in the following month and get bought back by the issuer within the next three months. The “control” group contains loans that become 30 days past due in the fourth month after securitization, transition to 60 days the following month and remain securitized in the subsequent three months. The strength of this empirical strategy is that one can plausibly assume that the fact that a loan became delinquent in the third versus the fourth month after being securitized is essentially random. This is PSV’s main argument for obtaining identification – to the extent that the precise timing of default is random, the securitization status of these loans is “as if” randomly assigned. We make clear below, however, that the fact that the authors also 7 There are two types of early-payment default (EPD) clauses, and it is important to distinguish between the two. EPD clauses were written into many of the loan sale agreements between mortgage originators and issuers of mortgage-backed-securities (MBS). In addition, EPD clauses were written into a small fraction of the Pooling and Servicing Agreements (PSAs) that govern the relationship between the issuer, servicer, and investors of MBS. An EPD clause in a loan sale agreement would require that the mortgage originator repurchase loans from the issuer if specific performance criteria were violated. On the other hand, an EPD clause in a PSA would require the issuer (or sponsor) of the securitization deal to buy back mortgages that became delinquent shortly after being securitized at the request of the investors. Although PSV refer throughout their paper to repurchased loans in the context of the relationship between the originator and the issuer, their analysis is concerned with the latter EPD clauses (those in PSAs) as they are the only EPD clause repurchases that are possible to identify in the LPS data. We will also refer to this type of EPD clause throughout our discussion. 8 PSV acknowledge in their paper that there is variation across deals in the exact timing of the EPD clauses (p. 383).

7

condition on whether the loans are actually repurchased or remain securitized invalidates what is, in principle, a promising identification strategy.

3.1. Implementation of EPD clauses as an instrument It is useful to be specific about the necessary conditions to employ EPD as an instrument and the precise way that EPDs are used in PSV. In its simplest form, the equation the authors estimate in the paper is Fi = β ∗ Si + η ∗ Xi + i

(1)

where Fi is an indicator for whether loan i is in foreclosure, Si is an indicator for whether the loan is in a “private-label” trust (i.e. it is equal to 1 if the loan is securitized), and Xi is a vector of control variables. The problem PSV attempt to address by using the EPD clauses is a standard omitted variables problem, i.e. that there may be one or more unobserved variables Zi that are correlated both with the outcome variable they measure and with the securitization status of the loan (i.e. Cov(Si , Zi ) 6= 0 and Cov(i , Zi ) 6= 0). For example, borrowers in securitized loan pools may be more subject to employment shocks that also drive the probability that a loan will be in foreclosure, but the econometrician cannot observe these shocks. This is equivalent to saying that the specification one would ideally want to estimate is Fi = β ∗ Si + η ∗ Xi + γ ∗ Zi + i

(2)

but this is not possible because Zi is not observable. Early-payment defaults, EP Di , would solve this problem provided three conditions were met, namely that 1. Cov(EP Di , Zi ) = 0 2. Cov(EP Di , i ) = 0 3. Cov(EP Di , Si ) 6= 0 where EP Di = 1 if loan i is securitized and misses a mortgage payment less than 90 days after securitization, and EP Di = 0 if delinquency happens after 90 days. If we restrict our attention only to securitized loans that become delinquent close to the 90 day threshold, the identifying assumption is that defaulting just before 90 days or just after that moment is essentially random. The sample is restricted to loans that start off in a securitization trust and, to the extent that EPDs require at least some loans to be repurchased (and thus 8

correlate with the ultimate securitization status of the loan), this should make EPDs an attractive instrument. The natural next step would be to implement a typical two-stage least squares estimator. In the first stage, EPD would be used as an instrument for whether the loan is securitized or not (for example, 3 months after delinquency), and in the second stage the instrumented securitization dummy could be used as an independent variable in a regression that predicts foreclosure. The instrumented securitization status would be given by SiInst = Si ∗

Cov(EP Di , Si ) V ar(EP Di )

(3)

and the correlation between SiInst and the unobserved variable Zi would be by assumption zero. The resulting estimated β coefficient would no longer suffer from omitted variables bias. In section 4 below, we implement this IV strategy and discuss the estimation results and their appropriate interpretation. In contrast, PSV do not use a two-stage estimator approach. Instead of creating the “predicted” securitization dummy SiInst , they create a new independent variable that can be expressed as 1. SiN ew = 0 if EP Di = 1 and Si,post−def ault = 0 2. SiN ew = 1 if EP Di = 0 and Si,post−def ault = 1 Crucially, this new variable conditions on the actual securitization status of the loan, not just on whether it defaulted before or after the 90-day threshold. Si,post−def ault is an indicator variable that takes a value of one if the mortgage was in a securitization trust after delinquency and zero otherwise. The first group (loans for which SiN ew = 0) are loans that missed a payment before the 90-day mark, subsequently defaulted (missed another payment to reach 60-days delinquent), and were then repurchased by the issuer. The second group (loans for which SiN ew = 1) contains loans that missed a payment after the 90-day mark (i.e. were not subject to the EPD clause), subsequently defaulted, and were then not repurchased, remaining securitized. The authors use this new variable as the independent variable in the foreclosure regression shown in equation (1). Sinew is not an instrument nor is it a predicted value from a first stage regression. Thus, the approach chosen by PSV is not an instrumental variables specification. PSV argue that the random timing of default implies that SiN ew is, by construction, a measure of securitization that is orthogonal to the omitted variables, i.e. Cov(SiN ew , Zi ) = 0. In fact, however, the random timing of default only ensures that Cov(EP Di , Zi ) = 0 whereas variation in SiN ew reflects both the EP Di variable and whether the investors choose to force 9

the issuer to repurchase the loan. For SiN ew to identify the effect of securitization on loan outcomes, it must satisfy the much stronger condition that: Cov(Si,post−def ault , Zi |EP Di ) = 0.

(4)

There are two ways that equation (4) might hold. First, if investors forced issuers to repurchase all loans that defaulted before the cutoff and no loans that defaulted after, then SiN ew = 0 = 1 − EP Di , and condition (1) (Cov(EP Di , Zi ) = 0) would ensure identification. However, the data decisively reject this possibility. The share of EPD loans repurchased is quite small. PSV find that approximately 9 percent of securitized loans that became delinquent within 90 days after being securitized are repurchased by banks within three months after the first instance of serious delinquency (see Figure 2 in PSV). We find a very similar percentage (approximately 11 percent) of repurchases in our sample of mortgages (see Figure 1 below). In either case, investors opt not to force the repurchase of approximately 90 percent of the loans that the EPD clause would render eligible for repurchase. One argument offered by PSV for the reason that investors force the repurchase of so few loans is that only a small fraction of securitization deals contained EPD clauses. However, since the LPS data that we use (and PSV uses) does not allow researchers to link loans with deals, there is no way to verify that there is any relationship between the clauses and the decision to force repurchase. Also, there is good reason to doubt that issuers only repurchase problem loans when required to by the contracts. Gorton and Souleles (2005) argue that issuers of asset-backed securities often provide such “state-contingent subsidies” even when accounting rules specifically forbid them from doing so. To make matters worse, even if one could show that all repurchases involve deals with EPD clauses, one would then have to argue that the EPD clauses were themselves randomly assigned to deals. Identification would fail if deals with EPD clauses differed in systematic and unobservable ways from deals without them.9 The second possible reason that equation (4) might hold is if investors force repurchases randomly. Unfortunately, there is no way to reliably tell if this is the case in the LPS data. PSV point out that the observable characteristics of loans that investors force issuers to repurchase are similar at origination to those that they do not. However, unobservable differences are the issue here, and therefore we view this observation as irrelevant to the validity of the identification assumption. 9 PSV find that the observable characteristics of mortgages in deals with EPD clauses are distinct at origination to those in deals that do not have such clauses (page 387), but that they have similar interest rates. While interest rates may in principle be a summary statistic for the riskiness of a loan (as PSV argue), the interest rate variable in this dataset is a very noisy measure of the true cost of a mortgage. So, even on observable characteristics, there seem to be differences between the two types of deals.

10

Ultimately, the “quasi-experiment” replaces one identification problem with another. The problem with the main specification of the paper (equation (1)) is that identification requires originators to randomly select the loans they sell to issuers. The “quasi-experiment” requires, on the other hand, that investors randomly select the loans to sell back to issuers. It is debatable which requirement, random initial sale or random repurchase, is more problematic. Arguably, information about the reason for early-payment default – job loss versus fraud, for example – is far more valuable than any information the issuer had at origination. PSV estimate alternative specifications of the test described above where they compare all loans that default before the 90-day cutoff and are repurchased to those that default in the same period and are bought back by originators. All of the critiques we discussed above apply to these alternative specifications, with the added concern that randomness of the timing of default is not used in the additional robustness tests.

4. An IV Strategy Using Early-Payment Defaults In this section we implement the IV estimation strategy described in Section 3.1 above in an attempt to obtain a true causal measure of the effect of securitization on the likelihood that a mortgage is foreclosed on or modified. Specifically, we use variation in the time of delinquency around the 90-day EPD threshold to instrument for whether or not a loan is securitized and then use the instrumented securitization variable in regressions where the likelihood of foreclosure and modification are the dependent variables of interest. According to the logic described above, a loan that becomes 30-days delinquent before the 90-day EPD threshold should have a higher probability of being repurchased by portfolio lenders compared to a loan that becomes 30-days delinquent after the threshold, and thus these loans will have a higher probability of being in portfolio at the time of serious delinquency (60+ days delinquent). This is confirmed in Figure 1. According to the figure, approximately 18 percent of loans that become delinquent in the first two months after securitization are repurchased. This percentage drops significantly to about 3 percent in the third month after securitization, which corresponds to the last month before the 90day EPD threshold. The percentage then drops to 1.3 percent in the fourth month, which is the month immediately after the threshold, and falls further to 0.5 percent in the fifth month after securitization. Hence, it is apparent that loans that first become delinquent before the threshold are more likely to be repurchased than those that become delinquent after the threshold. Thus, as long as the exact timing of delinquency is random around the 90-day threshold, and the repurchased mortgages that end up back on the originators’ balance sheets are representative of the population of portfolio-held loans, the IV estima11

tion strategy should produce unbiased estimates of the causal effect of securitization on the likelihood of renegotiation. We will first describe the data, then present the results of the IV estimation, and finally discuss whether these key assumptions are valid.

4.1. Data and Variable Construction One issue that must be confronted before implementing the IV estimation is how exactly to define the instrument. According to Figure 1 there is a significant drop in the repurchase fraction two months after securitization. Thus, the difference in the repurchase share for loans that become delinquent immediately before and after the threshold (third versus fourth month) is relatively small (3.0 versus 1.3 percent). PSV focus their analysis on these loans with the rationale that they are likely most similar in terms of both observable and unobservable characteristics, and thus whether a loan becomes delinquent in the third or fourth month can be considered random. However, due to the seemingly small difference in repurchase share between months three and four, we will construct our instrument using three different window sizes around the 90-day EPD threshold: A three-month window on either side of the threshold (loans that become delinquent in months 1, 2, and 3 versus loans that become delinquent in months 4, 5, and 6; a two-month window (loans that become delinquent in months 2 and 3 versus loans that become delinquent in months 4 and 5; and a 1-month window (loans that become delinquent in months 3 and 4). The pattern of repurchases displayed in Figure 1 suggests that the instrument will be stronger for the 2and 3-month windows compared to the 1-month window, but the drawback of using the larger windows is that the assumption that delinquency is random (i.e. uncorrelated with unobservable characteristics) is less likely to hold. Thus, the instrument is a dichotomous variable defined in the following way:

EP DIV,i =

 1 0

if T90,i − ∆t < M30DQ,i < T90,i if T90,i < M30DQ,i < T90,i + ∆t

where ∆t is the window size and takes on values of 1, 2, or 3 months, T90,i refers to the 90day EPD threshold that occurs between the 3rd and 4th month after securitization (i.e. T90,i takes a value between 3 and 4 in the above definition), and the term M30DQ,i corresponds to the month of first 30-day delinquency relative to the time of securitization. As mentioned above, our data come from LPS. The LPS data are collected from several of the largest U.S. mortgage servicers and cover a large fraction of active loans.10 The 10

The LPS loan-level dataset covers approximately 40 million active first lien mortgages and 8 million active second lien mortgages.

12

LPS data include a large number of standard mortgage underwriting fields. Loan-level attributes include borrower characteristics (e.g., origination FICO score, occupancy status, and documentation level), collateral characteristics (e.g., property type, original loan-tovalue ratio, and zip code), and loan characteristics (e.g., loan balance, lien holder type, and loan status). The monthly history of each loan appears in the data including their current payment/performance status. We use a sample of mortgages that is very similar (but not identical) to PSV. We provide a detailed list of our sample restrictions in Table 1. In Table 2 we display summary statistics for the three mortgage samples that correspond to the different windows that we use to construct the instrument, EP DIV,i . Within each sample, we break down the summary statistics by which side of the 90-day EPD threshold a mortgage experiences 30day delinquency, or equivalently by whether the mortgage is assigned a 1 or 0 value for the EP DIV,i variable. The mortgage characteristics displayed in the table are all included in the covariate sets of the regressions discussed below. These characteristics include the FICO score and loan-to-value ratio (LTV) at origination, the contract interest rate and size of the mortgage at origination, whether or not the mortgage is conforming or jumbo, the amortization schedule of the mortgage (interest-only or option-ARM which is a common type of negatively amortizing mortgage), the occupancy status of the borrower, whether the mortgage is a refinance or purchase, and whether the mortgage is subprime (where subprime is determined by the reporting servicer). There are no obvious discrepancies or patterns across the sample of mortgages that missed the first payment before the 90-day EPD threshold and the sample that missed the first payment after the threshold, for any of the window sizes. The average FICO score, LTV ratio, and contract interest rate appear to be slightly closer for the 1-month window sample compared to the 3-month window sample, but the averages are still quite close in the 3-month window sample. The average mortgage amount appears to be actually closer in the 3-month window sample, while the remainder of the characteristics look broadly similar across all of the samples. The last two rows of Table 2 display the foreclosure and modification fractions associated with each sample. Our foreclosure definition corresponds to completed foreclosure (i.e. when a property is either sold at auction to an arms-length buyer or a property is retained by the lender and marketed for sale), which is the same definition used by PSV. We measure whether a mortgage is modified using an algorithm that checks for changes in the contract terms of a mortgage that were not specified by the contract (i.e. interest rate changes, term changes, and principal balance changes).11 In this table as well 11

For details on the specifics of the modification algorithm and it’s precision, see Adelino, Gerardi, and Willen (2013).

13

as in all of the regressions, we adopt a 12-month horizon for foreclosures and modifications. That is, we measure whether a mortgage is foreclosed on or modified over a 12-month period from the month of first 60-day delinquency. According to Table 2, mortgages that become delinquent before the 90-day EPD threshold are about 2.5 percentage points more likely to experience foreclosure and about 1 percentage point less likely to be modified compared to loans that first become delinquent after the threshold. These differences appear to be insensitive to the window size around the threshold. To gain more insight regarding these differences in foreclosure and modification rates around the 90-day EPD threshold, we plot the fraction of foreclosures and modifications by month of first 30-day delinquency (relative to the month of securitization) in Figure 2. These are unconditional plots in the sense that they do not control for differences in observable characteristics. There are a few notable patterns in Figure 2. First, the foreclosure fraction appears to discretely drop by about 2 percentage points at the 90-day EPD threshold. After the threshold, the foreclosure fraction remains roughly constant around 20 percent. This pattern is consistent with the summary statistics presented in Table 2. This pattern is robust to controlling for differences in observable loan characteristics. Figure 4 displays the coefficient estimates associated with dummy variables that correspond to each month of 30-day delinquency relative to the time of securitization from a linear probability model where the probabilities of foreclosure and modification respectively are specified as the dependent variable and observable loan characteristics are included as covariates. The relationship between the incidence of foreclosure and the timing of delinquency is very similar to the pattern observed in Figure 2. Loans that miss a payment in the first 3 months after being securitized are more likely to experience foreclosure, controlling for observable loan characteristics. For the modifications, there appears to be a humped-shape relationship, whereby modification rates first rise in the time since securitization, until about 7 months and then slightly decrease. A similar pattern can be seen in the conditional modification rates displayed in Figure 4. It is certainly the case that modifications rates are lower for mortgages that become delinquent before the 90-day threshold, which is consistent with the summary statistics in Table 2, however, unlike the foreclosure plots, there does not seem to be a discrete increase around the threshold, but rather a steady upward trend. The simple summary statistics presented in this section show that loan repurchases are higher for delinquencies that occur before the 90-day EPD threshold, although the difference in repurchase rates is relatively small for delinquencies that occur in the months immediately surrounding the threshold (3 percent versus 1.3 percent). In addition, delinquencies that occur before the threshold are more likely to experience subsequent foreclosure and are also 14

more likely to receive a modification in the future compared to delinquencies that occur after the threshold. We now present results from the more formal IV analysis that we described above.

4.2. Estimation Results Table 3 displays the foreclosure estimation results from the IV regression that uses the time of first 30-day delinquency relative to the 90-day EPD threshold as an instrument for the securitization status of a mortgage.12 The first three columns in the table displays results from OLS, which in this context corresponds to a simple linear probability model. Results for three different samples are shown in each panel, with each sample corresponding to different window sizes around the 90-day EPD threshold (1-, 2-, and 3-month windows). The estimated OLS coefficients are negative and statistically significant with large magnitudes (in absolute value). Depending on the sample, securitized mortgages are between 14 and 16 percent less likely to experience foreclosure 12 months after the first 60-day delinquency, compared to loans that are not securitized. Column (4) displays results from the linear IV model where the instrument is defined over a 3-month window around the threshold (i.e. the instrument takes a value of 1 if the first 30-day delinquency occurred 1, 2, or 3 months after securitization, and takes a value of 0 if it occurred 4, 5, or 6 months after securitization). The results from the first stage indicate that the correlation between the instrument and securitization status is negative, as expected, and it is highly statistically significant. Loans that become delinquent before the 90-day threshold are approximately 4.5 percent less likely to be securitized at the time of serious delinquency. This is consistent with the pattern in Figure 1 that showed higher repurchase rates for delinquencies occurring before the threshold. The F-statistic from the first stage regression is significantly above the critical value obtained from the Stock-Yogo (2003) test for weak instruments using limited information maximum likelihood estimation (assuming a 10 percent size threshold). The results from the second stage indicate a strong, statistically significant negative effect of securitization on the incidence of foreclosure. According to the coefficient estimate, a private-label securitized mortgage is 24 percent less likely to experience a foreclosure in the 12 months after serious delinquency than a portfolio loan.13 The results in the fifth column of Table 3 that define the instrument over a 2-month window are similar. However, the results for the case in which we define the instrument over a 112

To define the securitization dummy, we use the reported securitization status of a loan at the end of the 12-month horizon for loans that survive for the entire horizon, and for loans that do not survive for the entire horizon, we use the securitization status of the loan in the last month that it was active. 13 In addition to controlling for the loan characteristics listed in Table 2, we also include year of origination fixed effects, MSA fixed effects, and quarter of serious delinquency fixed effects.

15

month window are much different. The first stage results appear to be much weaker, as delinquencies that occur in the third month after securitization are only 1.3 percent less likely to be securitized at the time of serious delinquency than delinquencies that occur in the fourth month after securitization. This effect is statistically significant however, and the Stock-Yogo weak IV test indicates that the instrument is not weak (although the test statistic drops significantly by an order of magnitude). However, the second stage results indicate an implausibly large effect, as the coefficient estimate is negative, but approximately 80 percent in absolute value. We interpret this as evidence that the linear probability model (LPM) may be inappropriate to use on this particular sample. Since we obtained implausibly large effects (in absolute value) for the 1 month window size using the LPM, we estimate a bivariate probit model in columns (7)–(9), which, unlike the LPM, restricts estimated probabilities to lie in the unit interval.14 The bivariate probit is estimated via maximimum likelihood estimation, and contains the same covariates as the LPM, including the instrument, which is included in the second equation of the system (the probability of securitization on whether the month of initial delinquency occured before the EPD threshold).15 The estimated average treatment effects (ATE) of securitization on the probability of foreclosure are significantly different from the local average treatment effects (LATE) that we obtained from the linear IV model.16 The ATEs are significantly smaller in absolute value than the LATEs obtained from the IV LPMs, for all window sizes. The ATEs for the 2- and 3-month window sizes are quite similar in magnitude to the results from the simple OLS models (about -17 percent). However, the ATE for the 1-month window size is much smaller in absolute value (-6 percent) and is not statistically different from zero. Thus, according to the bivariate probit results for the 2- and 3-month EPD window sizes, securitized loans are approximately 17 percent less likely to be foreclosed upon than portfolio loans. The dramatic differences in the estimated magnitudes of the LATEs and ATEs between the LPM and the bivariate probit model suggests that in this particular context, the fact that the LPM does not restrict the estimated probabilities to lie on the unit interval is a significant issue. These results are consistent with findings in Chiburis, Das, and Lokshin (2011), who show that when treatment probabilities (and outcome probabilities) are close to 0 or 1, the LATEs from linear probability models can be significantly biased 14

Note that in the context of a binary dependent variable and a binary endogenous regressor, the 2-stage probit, or IV Probit model as it is often called, yields inconsistent estimates. 15 We also estimated a simple probit model that does not address the potential endogeneity of the securitization decision for both foreclosure and modification outcomes. In both cases the marginal effects were the same sign and very similar in magnitude to the OLS results reported in columns (1) - (3) in Tables 3 and 4. 16 We used Stata code from Richard Chiburis’s website that was used in Chiburis, Das, and Lokshi (2011) to estimate the ATEs for the bivariate probit model, and obtained standard errors via bootstrapping.

16

away from the ATEs from bivariate probit models. Table 4 displays the modification estimation results. The structure of the table is identical to Table 3 but in this table the dependent variable corresponds to whether or not a loan is modified within 12 months of becoming delinquent, as opposed to foreclosed upon as in the previous table. Simple OLS estimation (columns 1 - 3), shows effectively zero correlation between securitized loans and modification frequency in the EPD sample. However, results from the linear IV and bivariate probit models tell a different story. The first stage results for the LPMs are identical to the case of foreclosure. The second stage results suggest that securitized loans are more likely to be modified than portfolio loans. The estimated average treatment effects for the 2 and 3-month windows are 0.142 and 0.172 respectively, which implies that securitized loans are more than 14 percent more likely to be modified than portfolio loans. The estimate for the 1-month window is significantly larger (32 percent), which is consistent with the pattern that we obtained in the foreclosure regressions. The bivariate probit estimation results are much smaller in magnitude, as the ATEs for the 2- and 3-month window sizes are 0.050 and 0.064, respecitively. This implies that securitized loans are about 5 percent more likely to be modified than portfolio loans. Also, as in the case of foreclosure, we see that the estimated ATE using the 1-month window is not statistically different from zero. Again, it appears as though the bivariate probit model is superior to the LPM in this context, as the results are much more stable across the window sizes, and are much more reasonable in terms of magnitudes. Based on the differences between the OLS and IV results, it is clear that selection is important in the context of modification decisions for the EPD mortgage sample. OLS estimation shows little correlation between securitization and modification, however, when EPD is used to instrument securitization, we find that securitized loans are significantly more likely to be modified. The negative estimates of the correlations in the residuals of the bivariate probits provide further evidence along these lines. For the 2- and 3-month windows, the correlation coefficient is -0.26 and -0.39 respectively, implies that mortgages that are more likely to be securitized are less likely to be modified in unobserved ways, so that when one controls for this, the impact of securitization on modification is positive.

4.3. Interpretation and Discussion In the previous section we showed that when early-payment defaults are used in a true instrumental variables setting, the impact of securitization on foreclosure is negative, and the impact of securitization on modification is positive. These results contradict those reported in PSV, who use early-payment defaults in the manner discussed in Section 3.1. In this section we discuss several reasons why we are skeptical of using EPD to identify 17

the causal impact of securitization on the incidence of foreclosure and modification. Our view is that both institutional and empirical evidence strongly suggests that when loans are repurchased from a securitization trust, they are treated very differently from portfolio loans that are never securitized. Thus, we are cautious to interpret the empirical results reported in Tables 3 and 4. However, for those that disagree and view EPD as a promising way of identifying the causal impact of securitization on foreclosure and modification, we have provided strong evidence that refutes the PSV hypothesis that frictions in the securitization process inhibit mortgage renegotiation. Our main concern with the EPD identification strategy is whether it is valid to extrapolate the results to the overall population of delinquent securitized and portfolio loans. There are two issues that make us skeptical of such an interpretation. First, an implicit assumption that must hold to make such an extrapolation valid, is that repurchased loans are treated similarly to other portfolio loans. PSV argue that they are, stating that “Once a loan has been sold back, the originators service these loans as any other loan on their balance sheets,” but they do not provide any concrete evidence that this is the case. Institutional evidence suggests the opposite, as issuers could unload repurchased problem loans in what was known as the Scratch & Dent (S&D) market. As described in an article in Mortgage Banker, the S&D market buys repurchased loans that institutions do not want to keep on their balance sheets: With nonperforming loans—those 60 days late on payment—investors are pushing back on the originators to keep those loans or take them back. In turn, originators are trying to get rid of them any way they can—so they are selling them to the S&D people at 30 and 40 cents on the dollar. But the buyers want to keep them and reform the loan, if possible (James Dowell, Mortgage Banking, September 17, 2007). Resale of problem loans is common in the LPS data. We find that approximately 60 percent of repurchased loans in the LPS data set end up in other securitization trusts. Since a large fraction of repurchased mortgages end up being serviced as private-label loans, the EPD identification strategy may not be very informative about portfolio loans in general. The second issue is the representativeness of borrowers who miss mortgage payments almost immediately after origination. Our sample is composed of borrowers that missed a payment within the first nine months of origination.17 These borrowers are likely different than borrowers that make a few years worth of payments before becoming delinquent in 17

Our sample is made up of borrowers that miss a payment within 6 months of securitization. Since we only keep mortgages that are securitized within 3 months of origination, this implies that our sample is made up of borrowers that missed a payment within 9 months of origination.

18

ways that are not captured by the variables in the LPS dataset. For example, there are many borrowers in the sample that either did not make a single mortgage payment or made only one payment. Many of these borrowers may have never even planned on moving into the properties, especially if they were investors trying to flip them quickly to other buyers. While these are not explicit critiques of the instrumental variables identification strategy for this specific sample of mortgages, they are issues that raise serious concerns about external validity.

5. Conclusion In this paper we develop and estimate an instrumental variables strategy for identifying the causal effect of securitization on the incidence of mortgage modification and foreclosure based on the early-payment default analysis performed by Piskorsi, Seru, and Vig (2010). We demonstrate the limits of the PSV “quasi experiment,” which we explicitly show is not an instrumental variables strategy, and thus does not use random variation in the probability that a loan is securitized. In particular, we argue that the PSV “quasi experiment” is just as likely to be plagued by endogeneity bias as a simple regression of foreclosure on securitization status. Our results from estimating a true IV regression based on early-payment defaults are in direct contrast to the results reported in PSV and Agarwal et al. (2011). We find that securitized mortgages are more likely to be modified and less likely to be foreclosed on by servicers. These results are consistent with the interpretation in Adelino, Gerardi, and Willen (2013) that low modification rates are not the result of contract frictions inherent in the mortgage securitization process. We are hesitant to place too much weight on the results from the IV regressions, however, as we have significant doubts regarding the validity of extrapolating the EPD results to the general population of delinquent mortgages. In addition, we view the institutional evidence behind the 90-day EPD threshold offered by PSV to be relatively weak. In short, we see little reason to place more emphasis on the EPD results than the empirical results found in the previous literature, which did not produce a compelling strategy to identify the causal effect of securitization on mortgage renegotiation.

19

6. References Adelino, M., Gerardi, K., P. Willen, 2013. “Why Don’t Lenders Renegotiate More Home Mortgages? Redefaults, Self-Cures and Securitization,” FRB Atlanta Working Paper. 200917. Agarwal, S., G. Amromin, I. Ben-David, S. Chomsisengphet, and D. D. Evanoff, 2011. “The Role of Securitization in Mortgage Renegotiation.” The Journal of Financial Economics 102(3): 559578. Angrist, J. D., 1990. “Lifetime earnings and the Vietnam era draft lottery: evidence from social security administrative records.” The American Economic Review, 80(3):313–336. Chiburis, R., Das, J., and M. Lokshin, 2011. “A Practical Comparison of the Bivariate Probit and Linear IV Estimators,” World Bank Policy Research Working Paper 5601. Dowell, J. “Servicing Scratch-and-Dent Loans Presents Special Challenges,” Mortgage Banking. September 2007. Gorton, G. and N. Souleles (2005). “Special Purpose Vehicles and Securitization,” NBER Working Paper 11190. Keys, B. J., Mukherjee, T., Seru, A., and V. Vig, (2010). “Did securitization lead to lax screening? Evidence from subprime loans,” The Quarterly Journal of Economics 125(1): 307–362. Krainer J. and E. Laderman (2009). “Mortgage Loan Securitization and Relative Loan Performance,” Federal Reserve Bank of San Francisco Working Paper 2009-22. Mian A. and A. Sufi (2009). “The consequences of mortgage credit expansion: Evidence from the US mortgage default crisis,” The Quarterly Journal of Economics 124(4): 1449– 1496. Piskorski, T., A. Seru, and V. Vig (2010). “Securitization and distressed loan renegotiation: Evidence from the subprime mortgage crisis,” Journal of Financial Economics 97(3): 360– 397.

20

Figure 1. Early-payment Default Repurchases

20%

18%

16%

21

% of Loand Repurchased

14%

12%

10%

8%

6%

4%

2%

0% 1

2

3

4

5

6

7

8

9

10

11

12

Month of 30!day Delinquency (Relative to Month of Securitization)

Notes: This figure plots the fraction of loans that are repurchased as a function of the first time they become 30-days delinquent after securitization. We impose the restriction that the repurchase must occur within 6 months of the time of first delinquency. The black vertical line corresponds to the 90-day threshold that PSV argue is stipulated in many (but not all) Pooling and Service Agreements. Thus, loans that become delinquent before this threshold should be more likely to be repurchased than loans that become delinquent after the threshold. This is confirmed in the figure, although there is no sharp discontinuity one each side of the threshold. Rather there appears to be a sharp discontinuity 60 days after securitization, which PSV also find in their sample (See Figure 2 in their paper).

Figure 2. Unconditional Incidence of Foreclosure and Modification by Month of 30-day Delinquency Foreclosure

24%

23%

% Fo oreclosured (12!month Horizon)

22%

21%

20%

19%

18%

17%

16%

15% 1

2

3

4

5

6

7

8

9

10

11

12

9

10

11

12

Months between 30DQ and Securitization

Modification

10%

% Modified (12!month Horizon)

9%

8%

7%

6%

5%

4%

3% 1

2

3

4

5

6

7

8

Months between 30DQ and Securitization

Notes: Figures are plots of the foreclosure and modification fractions respectively by month of first 30day delinquency (relative to the month of securitization). Foreclosure is defined to be the completion of foreclosure proceedings by the servicer within 12 months of 60-day delinquency (sale to an arms-length buyer at auction or the retention of the property by the bank). Modification is defined to be changes in the terms of a loan that are not stipulated in the original contract that occur within 12 months of 60-day 22 delinquency. These are raw fractions and do not control for observable loan or borrower characteristics.

Figure 3. Different approaches to identifying the effect of securitization. (1) Using a Dummy Military Service?

Securitized? Yes

No

Yes

No

Treatment

Control

Treatment

Control

(2) Instrumental Variables Securitized?

Military Service?

No (Si = 0)

Treatment

Yes (EP Di = 1)

Control

EPD?

No (EP Di = 0)

Yes

High Draft Number?

Yes (Si = 1)

No

No

Treatment

Yes

Control

(3) PSV “Quasi-Experiment” Securitized?

Treatment

EPD?

No (EP Di = 0)

Yes (EP Di = 1)

Military Service?

No (Si = 0)

Control

Yes

High Draft Number?

Yes (Si = 1)

23

No

Yes

No

Treatment

Control

Figure 4. Conditional Incidence of Foreclosure and Modification by Month of 30-day Delinquency Foreclosure

4.0%

Month of 3 30 DQ Dummy Estimates !! Foreclosure

3.0%

2.0%

1.0%

0.0%

1.0%

2.0%

3.0%

4.0% 1

2

3

4

5

6

7

8

9

10

11

8

9

10

11

Month of 30 DQ (since securitization)

Modification

2.0%

Month of 30 0 DQ Dummy Estimates !! M Modification

1.5% 1 5%

1.0%

0.5%

0.0%

0.5%

1.0%

1.5% 1

2

3

4

5

6

7

Notes: These figure displays coefficient estimates associated with dummy variables that correspond to each month of 30-day delinquency relative to the time of securitization from a linear probability model where the probabilities of foreclosure and modification respectively are specified as the dependent variable and observable loan characteristics are included as covariates.

24

Table 1

Sample Restrictions

1.

Loans originated in between January 1, 2005 and June 30, 2007.

2.

Mortgage payment history through August 31, 2008.

3.

First-lien mortgages only.

4.

Drop observations with incomplete information about original credit scores, original interest rates, origination amounts, and property values.

5.

Drop mortgages originated outside of Metropolitan Statistical Areas.

6.

Keep only mortgages with maturities of 15, 20, and 30 years.

7.

Keep only mortgages that entered the LPS database within four months of the origination date.

8.

Drop loans with FICO scores less than 500 and greater than 850.

9.

Drop loans with LTV greater than 150 percent.

10. Drop loans originated in Alaska, Hawaii, and other noncontinental areas. 11. Drop loans that had their servicing rights transferred to a servicer outside of LPS’s coverage. 12. Keep loans that we either see transition from portfolio to securitized (private-label), or that are reported as securitized during the month of origination. 13. Drop loans that become securitized more than 6 months after origination. Notes: This table displays the restrictions applied in obtaining the estimation sample. These restrictions are very similar, but not identical, to the restrictions used by Piskorski, Seru, and Vig (2010). The major difference is that we include loans originated in the first half of 2007 and follow loan performance through August 2008, while PSV exclude 2007 originations and follow loan performance through March 2008.

25

Table 2

Summary Statistics

26

# Mortgages Fico LTV (%) Interest rate Mortgage amount ($ thousands) Jumbo (d) Interest-only (d) Option-ARM (d) Non owner-occupant (d) Refinance (d) Subprime (d) Securitized (d) Foreclose (d) Modify (d)

Window Around 90-day Threshold 3 months 2 months 1 month EP DIV,i = 1 EP DIV,i = 0 EP DIV,i = 1 EP DIV,i = 0 EP DIV,i = 1 EP DIV,i = 0 44,798 54,605 38,644 38,006 21,856 19,870 614 619 615 617 617 615 80.06 79.80 79.97 79.80 79.92 79.93 8.22 8.05 8.19 8.09 8.17 8.14 237.7 240.4 239.7 237.5 240.9 233.9 0.15 0.15 0.15 0.15 0.16 0.14 0.18 0.20 0.19 0.20 0.19 0.19 0.22 0.21 0.24 0.21 0.22 0.21 0.09 0.07 0.09 0.07 0.09 0.07 0.49 0.52 0.49 0.51 0.48 0.51 0.73 0.71 0.73 0.72 0.72 0.73 0.94 0.99 0.95 0.99 0.97 0.99 0.227 0.202 0.226 0.201 0.229 0.207 0.076 0.088 0.077 0.088 0.078 0.087

Notes: This table reports sample averages of the dependent variables and main regressors included in our IV estimation. The averages are broken down by whether the first missed payment occurred before or after the 90-day EPD threshold, and are calculated for window sizes of 1, 2, and 3 months. The samples include only loans that missed a mortgage payment within 6 months of securitization, and missed a second payment at a subsequent point in time. Foreclose is an indicator variable that takes a value of 1 if the lender completed foreclosure proceedings within 12 months of the first 60-day delinquency, and modify is an indicator variable that takes a value of 1 if the mortgage received a modification within 12 months.

Table 3

Estimation Results: Effect of Securitization on Incidence of Foreclosure (1) Estimation method Window around 90-day threshold Securitized (PLS indicator)

3 months

(2) OLS 2 months

(3) 1 month

-0.163*** -0.146*** -0.139*** (0.0149) (0.0136) (0.0188)

(4)

(5) Linear IV 3 months 2 months -0.238** (0.0538)

(6) 1 month

-0.307*** -0.789*** (0.0730) (0.2922)

(7)

(8) (9) Bivariate Probit 3 months 2 months 1 month -0.174** (0.0369)

-0.175*** (0.0466)

-0.060 (0.0808)

. . . . 0.049 (0.0518) N Y Y Y MSA 99,403

. . . . 0.071 (0.0678) N Y Y Y MSA 76,650

. . . . -0.085 (0.1316) N Y Y Y MSA 41,726

First Stage:

27

Delinquent before 90 days (EPD indicator)

. .

. .

. .

F-Statistic for endogeneity (Wu-Hausman) Critical Value: χ2 (1) (2-sided 5% test) F-Statistic for weak IV test Stock-Yogo critical value (10% LIML size) ρ – correlation between residuals

. . . . . . Y Y Y Y MSA 99,403

. . . . . . Y Y Y Y MSA 76,650

. . . . . . Y Y Y Y MSA 41,726

MSA fixed effects Quarter of serious delinquency effects Year of origination effects Other controls Clustering unit # Mortgages

-0.045*** -0.037*** -0.013*** (0.0011) (0.0012) (0.0014)

19.55 5.024 1569 16.38 . . Y Y Y Y MSA 99,403

17.67 5.024 974 16.38 . . Y Y Y Y MSA 76,650

7.62 5.024 90.67 16.38 . . Y Y Y Y MSA 41,726

Notes: This table reports results of the estimated effects of securitization on foreclosure. The dependent variable is indicator variable that takes a value of 1 if the loan completes the foreclosure process within 12 months of the first 60-day delinquency. The regressor of interest is “PLS,” which is an indicator variable for whether the loan was privately securitized (as opposed to being on a bank’s balance sheet) 12 months after the first 60-day delinquency (or the last month that a mortgage is active in the LPS dataset for loans that do not survive for the full 12 months). Columns (1)–(3) display coefficient estimates from simple OLS regressions, which correspond to linear probability models in this context. Columns (4)–(6) display local average treatment effects (LATE) from a linear probability model where the securitization status of a loan is instrumented by the month in which the first missed mortgage payment occurred relative to the EPD 90-day threshold. Columns (7)–(9) display average treatment effects (ATE) from a bivariate probit with the same instrument used to identify the effect of securitization. In addition to the fixed effects displayed in the table, other control variables include the FICO score and loan-to-value ratio (LTV) at origination, the contract interest rate and size of the mortgage at origination, whether or not the mortgage is conforming or jumbo, the amortization schedule of the mortgage (interest-only or option-ARM which is a common type of negatively amortizing mortgage), the occupancy status of the borrower, whether the mortgage is a refinance or purchase, and whether the mortgage is subprime (where subprime is determined by the reporting servicer). T-statistics are reported below the marginal effects, and standard errors are clustered at the MSA level.

∗ , ∗∗ ,

and

∗∗∗

denote statistical significance at the 10, 5, and 1 percent levels, respectively.

Table 4

Estimation Results: Effect of Securitization on Incidence of Modification (1)

(2) (3) OLS 3 months 2 months 1 month

(5) Linear IV 3 months 2 months

1 month

(8) (9) Bivariate Probit 3 months 2 months 1 month

-0.003 (0.0063)

-0.006 (0.0069)

-0.015 (0.0101)

0.142*** (0.0390)

0.324 (0.2092)

0.050*** (0.0117)

Delinquent before 90 days (EPD indicator)

. .

. .

. .

F-Statistic for endogeneity (Wu-Hausman) Critical Value: χ2 (1) (2-sided 5% test) F-Statistic for weak IV test Stock-Yogo critical value (10% LIML size) ρ – correlation between residuals

. . . . . . Y Y Y Y MSA 99,403

. . . . . . Y Y Y Y MSA 76,930

. . . . . . Y Y Y Y MSA 41,726

Estimation method Window around 90-day threshold Securitized (PLS indicator)

(4)

0.172*** (0.0533)

(6)

(7)

0.064*** (0.0088)

0.012 (0.0695)

. . . . . . . . -0.256*** -0.388*** (0.0775) (0.0887) N N Y Y Y Y Y Y MSA MSA 99,403 76,930

. . . . -0.105 (0.2132) N Y Y Y MSA 41,726

First Stage:

28

MSA fixed effects Quarter of serious delinquency effects Year of origination effects Other controls Clustering unit # Mortgages

-0.045*** -0.037*** -0.013*** (0.0011) (0.0012) (0.0014)

13.41 5.024 1569 16.38 . . Y Y Y Y MSA 99,403

10.54 5.024 974 16.38 . . Y Y Y Y MSA 76,930

2.45 5.024 90.67 16.38 . . Y Y Y Y MSA 41,726

Notes: This table reports results of the estimated effects of securitization on modification. The dependent variable is indicator variable that takes a value of 1 if the loan receives a modification within 12 months of the first 60-day delinquency. The regressor of interest is “PLS,” which is an indicator variable for whether the loan was privately securitized (as opposed to being on a bank’s balance sheet) 12 months after the first 60-day delinquency (or the last month that a mortgage is active in the LPS dataset for loans that do not survive for the full 12 months). Columns (1)–(3) display coefficient estimates from simple OLS regressions, which correspond to linear probability models in this context. Columns (4)–(6) display local average treatment effects (LATE) from a linear probability model where the securitization status of a loan is instrumented by the month in which the first missed mortgage payment occurred relative to the EPD 90-day threshold. Columns (7)–(9) display average treatment effects (ATE) from a bivariate probit with the same instrument used to identify the effect of securitization. In addition to the fixed effects displayed in the table, other control variables include the FICO score and loan-to-value ratio (LTV) at origination, the contract interest rate and size of the mortgage at origination, whether or not the mortgage is conforming or jumbo, the amortization schedule of the mortgage (interest-only or option-ARM which is a common type of negatively amortizing mortgage), the occupancy status of the borrower, whether the mortgage is a refinance or purchase, and whether the mortgage is subprime (where subprime is determined by the reporting servicer). T-statistics are reported below the marginal effects, and standard errors are clustered at the MSA level.

∗ , ∗∗ ,

and

∗∗∗

denote statistical significance at the 10, 5, and 1 percent levels, respectively.