Regressive Mortgage Credit Redistribution in the Post-Crisis Era

Abstract

“Reducing onerous and unnecessary origination and servicing requirements (there are 3,000 federal and state requirements today) and opening up the securitization markets for safe loans would dramatically improve the cost and availability of mortgages to consumers […] And these would not be subprime mortgages but mortgages that we should be making.”

—Jamie Dimon, Letter to Shareholders, 2018

To understand the drivers of the 2008-2009 financial crisis, a large literature studies the aggregate dynamics of mortgage origination in the United States in the period between 2001 and 2007.¹ Moreover, a nascent literature assesses the effects of specific regulatory interventions during and after the financial crisis on mortgage origination by U.S. lenders.² So far, less attention was devoted to studying the aggregate dynamics of the U.S. mortgage market in the post-crisis period,³ despite the fact that these dynamics are likely to affect disparate outcomes, ranging from the emergence of entrepreneurship and economic growth (see, e.g., Schmalz et al. 2017) to households’ ability to accumulate wealth over time (see, e.g., Andersen et al.  2020).

In this paper, we document and study a sizable redistribution of mortgage credit from low- and midsize loans to large loans, which started in 2011 and has increased since. We do so using the quasi-universe of approved and rejected mortgage applications for both traditional and FinTech lenders in the United States over the period 2008-2017. Our analysis suggests that the supply side of the market drove this redistribution, rather than changes in borrowers’ characteristics, and is motivated by the plot below.

Open in new tabDownload slide

The left panel plots the loan-size distribution of the mortgages originated by Bank of America, one of the top-five lenders throughout our sample period. The mortgage origination dynamics are similar for all other large U.S. originators with the exception of Quicken Loans, which became a large originator towards the end of our sample period, in 2015.⁴ The overall distribution of credit by loan size started to shift to the right in 2011, and has kept moving to the right since. For large lenders, the bunching at the standard conforming loan limit of |${\$}$|417K, which was pronounced in 2008, has also dropped and fully disappeared by 2017. These dynamics imply that the incentives to bunch loans just below the conforming loan limit—which provides lenders with a guarantee from GSEs—instead of originating slightly larger nonconforming loans has completely disappeared for the largest originators.⁵ The right panel plots the distribution for lenders outside the top five. The distribution of mortgage credit has shifted to the right since 2011 also for smaller lenders, including all the FinTech lenders that report based on HMDA, although to a lesser extent. Importantly, smaller lenders increased the average size of their originated loans within the baseline conforming loan limit, but barely originate loans above the limit.⁶

Building on these motivating plots, in the first part of the paper we show that the redistribution of lending from small and medium-size loans to large loans holds for the number, size, and approval rates of loans for all lenders (first fact), and especially for the largest lenders (second fact). Moreover, the extent of redistribution increased proportionally with the size of the lender (third fact). The fourth fact we document is that loans further away from the conforming loan limit(s) from above and below drive this redistribution. This fact suggests that systematic differences in the incentives to originate conforming or jumbo loans cannot explain our results in full. Otherwise, the redistribution would be driven by loans close to the conforming loan limit(s) in both directions.

The baseline facts do not allow us to disentangle the roles of the supply and demand sides of origination, because large and small lenders are not matched randomly to households demanding credit. To tackle this concern, in the second part of the paper we analyze the extensive margin of mortgage origination—applications’ approval rates based on applicant and loan characteristics. Because our data cover the quasi-universe of approved and rejected mortgage applications, this analysis keeps constant the overall demand for mortgage credit at each point in time (total applications). We observe a set of demand characteristics: applicants’ race and reported income, the local racial composition, average household income, local house prices, and the local share of foreclosed properties. Even after controlling for these features of mortgage demand, we find that since 2011 larger loans are more likely to be approved than smaller loans, and this pattern is stronger for large lenders relative to other lenders. By estimating quantile regressions, we show that the mortgage-credit redistribution from small and medium-sized loans to large loans holds also at the intensive margin, that is, the dollar value of approved loans over time and across lenders.

All results are similar if we further restrict the variation within county-year pairs to absorb local business cycle shocks, if we control for the concentration of local lending markets (Scharfstein and Sunderam  2016), and if we exclude nonbank mortgage originators, which suggests substitution between bank and nonbank lenders does not explain our results (Buchak et al.  2018). Results are also similar if we compare areas and periods with different impacts of unconventional monetary policy measures (see, e.g., DiMaggio, Kermani, and Palmer  2016; Chakraborty, Goldstein, and MacKinlay  2020)

The demand for mortgages might still drive our results if unobservable borrower characteristics had discouraged middle-class households from applying for mortgages, and if discouragement was higher for prospective applicants to large lenders. We thus propose an instrumental variable strategy to obtain quasi-exogenous variation in the share of large lenders’ county-level mortgage origination activity, which is a proxy for the size of lenders households can access locally. We instrument the share of large lenders active in each county-year from 2008 to 2017 with the share of mortgage originations by large lenders in the county in 2007, before the 2008-2009 financial crisis.

The rationale for this instrument is that the share of large lenders in a county before the financial crisis is not determined by (unforeseeable) financial-crisis-related shocks, or by the fiscal and monetary policy measures implemented after the crisis. Because of the costs of relocating branches as well as regulation-driven geographic segmentation (Berger, Demsetz, and Strahan  1999), inertia occurs in the local supply structure of financial services, which makes the instrument relevant. Similar strategies based on the heterogeneity of exposure to supply-side shocks have been used in recent economic research (see, e.g., Mian and Sufi  2012; Chodorow-Reich  2014; Agarwal et al.  2017).

The identifying assumption is that the variation in the share of large lenders in 2007 does not affect the amounts lent since 2011 through channels different from the share of large lenders since 2011. This exclusion restriction cannot be tested, and hence we propose evidence to assess its plausibility. First, we document that the evolution of business cycle indicators at the county level, such as gross domestic product (GDP) per capita, unemployment rates, labor force participation and hours worked, share of homeowners, average housing rents, and the yearly growth of local bank deposits before and after the financial crisis was similar across counties with a high or low share of large lenders in 2007. We detect no pre-trends in the economic characteristics of counties treated with different shares of large lenders. Nor did counties with different shares of large lenders behaved differently after the financial crisis, suggesting that the policy measures implemented during the Great Recession did not affect these counties differently.⁷ Second, following Chodorow-Reich (2014), we show the outcome variable and covariates are balanced across counties sorted by our instrument. The instrumental variable analysis confirms our baseline results in terms of both statistical significance and magnitude.

In the last part of the paper, we assess the extent to which a set of nonexclusive supply-side channels are consistent with one or more of the trends we document, in the spirit of Chen, Hanson, and Stein’s (2017) discussion of small business lending. We base our assessment on direct and/or indirect arguments and evidence about the broader economic implications each channel implies. With this exercise, we aim to provide guidance about which channels are more or less promising for tailored, but possibly more narrow and less externally valid, causal tests.

We organize the analysis of channels along two dimensions: Risk-based channels—changes in the risk profile of loans—and regulation-based channels—changes in mortgage-market regulations that might have reduced the relative incentives to originate smaller loans, especially for larger lenders. Note that some of the channels may belong (at least in part) to both categories.

Starting with risk-based channels, we first ask whether banks’ risk management concerns might explain our results. Lenders may have started to originate large loans to decrease the riskiness of their assets in an effort to comply with stricter capital requirements after the financial crisis. This explanation implies that riskier banks drive our results, but our facts are not replicated if we use direct proxies for bank-level risk instead of size in our analysis.

Second, we consider the fact that GSEs have tightened their “putback” policy after the financial crisis: They imposed stricter rules for the buy back of loans whose origination documents were found to misrepresent borrowers’ ability to repay.⁸ This stricter policy created uncertainty in the effectiveness of the GSEs’ guarantee, and hence might have dis-incentivized lenders to originate conforming loans. We show that the timing of the relaxation of this policy is inconsistent with our results, because the redistribution we document became starker after GSEs relaxed this policy relative to when the policy was stricter.

Third, we consider the freeze of the private-label securitization market since 2008.⁹ Jumbo loans were typically securitized privately before the crisis, but this option has largely disappeared since the crisis. The residential-mortgage private-label securitization market remained stagnant throughout our sample, not only between 2008 and 2010 but also between 2011 and 2017, and, hence, it cannot have determined differential incentives of lenders across these two periods.

Fourth, because we observe no discontinuities in the redistribution between small dollar loans (e.g., loans below |${\$}$|70K)—which had a high risk profile before the crisis—and conforming loans of slightly larger size, stricter standards in the origination of small-dollar loans after the crisis are unlikely to explain our results either.

Moving on to regulation-based potential channels, we start by considering lenders’ reaction to subsequent waves of unconventional monetary policy implemented during the Great Recession. Chakraborty, Goldstein, and MacKinlay (2020) show that MBS repurchase waves by the Federal Reserve created cycles in banks’ lending behavior. To assess this potential explanation, we repeat our analysis separately for the years following fourth-quarter MBS purchases—when the policy was most effective—relative to other years. Consistent with Chakraborty, Goldstein, and MacKinlay (2020), MBS repurchases increased the overall amount of mortgage lending. But, the redistribution across loan sizes we document, if anything, was larger in the years in which the policy was least effective.

A second set of regulation changes has affected the costs of mortgage origination. We consider two pieces of regulation—the Temporary Payroll Tax Cut Continuation Act (TCCA) and the Dodd-Frank Wall Street Reform and Consumer Protection Act (Dodd-Frank).

TCCA determined a series of increases of the annual insurance premiums lenders paid to GSEs (“g-fees”).¹⁰ The increases caused a higher per-loan cost of originating conforming loans. An implication of this channel is that the redistribution effect should be concentrated just above the conforming loan limit, because lenders’ incentives to bunch origination at the limit instead of originating otherwise slightly larger mortgages had dropped. This implication of the g-fee channel is the opposite of what we document in our fourth fact.

The approval of Dodd-Frank has also increased both the expected fixed costs and per-loan costs of mortgage origination. For instance, lenders were required to establish an internal training system and to provide special training to all loan officers at their branches, which increased the fixed costs of producing loans of any size. Higher fixed costs of origination decreased lenders’ incentives to originate all mortgages, but small mortgages should have been affected more than large mortgages, because the expected per-mortgage profit decreased relatively more for small mortgages relative to large mortgages. Dodd-Frank has also imposed a thorough yet costly verification of the income customers reported at the time of the application. This procedure increased the per-loan costs of origination. Because the income verification procedure carries the same cost, regardless of the size of the loan, it also decreased the per-loan profit of smaller loans by more relative to the per-loan profit of larger loans. All banks thus had the incentive to originate larger loans and reduce their origination of small and medium-sized loans, consistent with the first and fourth facts we document.

The cost channel is also consistent with the second and third facts, which state that the incentives to originate larger mortgages are higher the larger the size of the lender. This is because originating large (nonconforming) loans is relatively easy for large lenders, but more difficult for smaller lenders, for at least three reasons. First, lenders have to keep jumbo loans on their balance sheets due to the frozen private-label securitization market. Small lenders with limited investment capital would have undiversified investments if they originated a small number of large loans. Second, larger banks can engage in cross-selling to wealthy customers more than small banks, credit unions, and other originators, such as FinTech lenders, which barely provide any services other than loan origination. Third, large lenders have geographically diversified operations and can redirect their origination activity to areas in which a higher supply of large loans can be absorbed.

Anecdotally, financial institutions reacted immediately after the approval of Dodd-Frank, even though several provisions were not self-executing, because the approval affected lenders’ expectations about per-loan profits and hence their willingness to engage in mortgage origination, relative to other business segments.¹¹ This behavior is not surprising, because lenders had to invest substantial resources up-front, such as building new training infrastructures and hiring specialized employees to execute the in-depth income verification procedures, in order to comply with the new provisions at the time of their execution.

1. Related Literature

This paper contributes to two strands of the economics and finance literature. First, we fit in the upcoming literature that studies the mortgage-origination behavior of financial institutions. Most of this literature focuses on origination before the 2008-2009 financial crisis, see Mian and Sufi (2009, 2016), Guiso, Sapienza, and Zingales (2013), Agarwal et al. (2014), Palmer (2015), Adelino, Schoar, and Severino (2016), Albanesi, DeGiorgi, and Nosal (2016), and Foote, Loewenstein, and Willen (2016). Our paper focuses on lenders’ mortgage origination behavior after the crisis. Chen, Hanson, and Stein (2017) describe the lending behavior of the four largest U.S. banks towards small businesses after the crisis. Lending to small businesses by large banks has decreased in the years immediately after the crisis (2008-2010) and has remained depressed between 2011 and 2014. The redistribution of consumer mortgage lending we document has started after 2011, and has not stopped until 2017, the last year in our sample. In both cases, higher regulatory compliance costs introduced during the crisis are, at least in part, likely drivers of these dynamics.¹²

Our analysis also suggests that well-known facts regarding mortgage-originating behavior might have changed permanently after the crisis. For instance, for the largest lenders, the bunching at the conforming loan limit, a robust fact that was used to estimate the elasticity of house prices to interest rates (DeFusco and Paciorek  2017), might have disappeared for good.

Second, we contribute to the line of research that studies the effects of fiscal and monetary policy on households’ consumption and saving choices, such as fiscal stimulus and unconventional monetary policy (see, e.g., Mian and Sufi  2012, Green et al.  2014; Broda and Parker  2014). DiMaggio, Kermani, and Palmer (2016) and Rodnyansky and Darmouni (2016) study the effects of unconventional monetary policy—the long-duration large-scale asset purchase programs (LSAPs)—on lending behavior in the period 2008-2013. Chakraborty, Goldstein, and MacKinlay (2020) document an increase in lending after MBS repurchase waves by the Federal Reserve. We confirm this aggregate increase in lending, but we find that the redistribution we document was, if anything, stronger in the years not preceded by fourth-quarter MBS repurchases. On the fiscal policy side, Agarwal et al. (2017) find large banks are much less likely to engage in renegotiation after the program, so our baseline estimates are, if anything, a lower bound of the true effect of bank size on origination. Moreover, because Agarwal et al. (2017) find higher exposure to HAMP resulted in lower house price declines and lower foreclosure rates, we directly control for these quantities in our baseline or robustness specifications. Another measure of fiscal policy that targeted the mortgage market after the crisis was the First-Time Homebuyer Credit (FTHC) (see, e.g., Berger, Turner, and Zwick  2020; Brogaard and Roshak  2011; Hembre  2015).

2. Data

Our main source of data is the Home Mortgage Disclosure Act (HMDA) data set for the years 2007-2017, which we obtain through the Consumer Financial Protection Bureau (CFPB). The data set contains the vast majority of mortgage applications over the sample period.¹³ For each mortgage application, we collect the application amount, the loan status (approved/rejected), the lien of the loan, the purpose of the loan (home purchase/refinancing/home improvement), the owner occupancy status, the lender identifier, the applicant’s income, the race and ethnicity of the applicant, and the location of the applicant (county/census tract). For all individual-loan results, the working sample uses only loan applications for home-occupied new purchases secured by a first lien. Our preferred definition of rejected loans includes only HMDA action code 3, that is, “denied applications.”¹⁴

Our second source of data is Zillow, through which we obtain the time-series of house prices and the number of houses foreclosed for every county-year. Because individual ZIP codes may contain more than one census tract, and because individual census tracts may belong to more than one ZIP code, we follow Adelino, Schoar, and Severino (2016) and use the Missouri Census Data Center bridge to aggregate our individual-loan data at the ZIP code level. Because the census tracts definitions vary over time, we use the two alternative bridges available from the Missouri Census Data Center for the periods 2007–2011 and 2012–2017.¹⁵

For the parallel trends analysis associated with our instrumental variable strategy, we obtain information on a set of county level macroeconomic variables over time from a set of sources. Specifically, we obtain:

• Median household income at the county level from the U.S. Department of Housing and Urban Development (HUD);

• Household-level income, labor force participation, number of weekly hours worked, homeownership status, and monthly rent paid for a representative 5% of the U.S. population from the Annual Community Survey (ACS), maintained by IPUMS;

• County-level unemployment rates from the U.S. Bureau of Labor Statistics;

• County-level GDP per full-time units (FTU) from the U.S. Bureau of Economic Analysis;

• Dollar value of county-level bank deposits from the U.S. Census Bureau.

Table 1 reports the summary statistics for the main covariates and outcomes. Panel A refers to the sample of approved loans. Our broader sample includes 26,419,090 individual-loan applications that were approved from 2007 to 2017. The median amount lent is |${\$}$|186K, but the loan amounts vary widely from |${\$}$|64K at the 5th percentile to |${\$}$|516K at the 95th percentile. On average, 14% of the mortgages are originated by the top-five U.S. lenders. The median reported gross income of approved applicants is |${\$}$|71K, and the variation in income is similar to the variation in requested loan amounts. On average, 8% of the applicants are Black, 6% are Asian, and 10% are Latino. The median house price in the counties of approved loans is |${\$}$|178.3K.

Panel B of Table 1 reports the corresponding descriptive statistics for the 4,954,288 rejected applications. The median loan amount requested by rejected applicants (⁠|${\$}$|145K) is lower than the one requested by approved applicants, and this median masks a distribution with fatter tails. A similar comparison holds for the distribution of rejected applicants’ income and of median house prices in the counties in which applicants are rejected. The share of rejections by top-five U.S. lenders—compared to other lenders—is higher than the share of approvals. Moreover, rejected applicants are more likely than approved applicants to be Black (15%) and Latino (15%), but similarly likely to be Asian (6.0%).

A measurement issue that has garnered considerable attention in the mortgage origination literature is that households’ gross income is self-reported in HMDA, and hence overstated. We believe overstatement of income may be less of an issue in our sample, because of the put-back policies enacted by GSE’s starting in 2008. Moreover, our aim is not to study the relation between income and mortgage origination as earlier research did. In our setting, income proxies for the demand for mortgages across households.

3. Mortgage Origination after the Financial Crisis: Raw Data

In this section, we introduce four novel facts about U.S. lenders’ mortgage origination behavior in the post-crisis period, from 2008 to 2017. The facts represent secular trends that have started in 2011 and have not reverted until the end of our sample period.

3.1 Fact 1: The size of originated mortgages has increased since 2011

Since 2011, U.S. lenders have started to approve relatively larger loans over time, whereas they have been cutting back on small and medium-sized loans over time.

Figure 1 summarizes the dynamics of aggregate mortgage origination between 2008 and 2017. In panel A, we compute the yearly average number of loans originated in each U.S. county across five loan-size bins. The loan-size bins are below |${\$}$|200K, between |${\$}$|201K-|${\$}$|400K, |${\$}$|401K-|${\$}$|600K, |${\$}$|601K-|${\$}$|800K, and above |${\$}$|800K. For each county, we compute the number of loans approved in the size bin regardless of the type of lender that originated the loan. Starting from the far left plot, we can see that the average number of loans approved for values up to |${\$}$|200K has dropped over time, ranging from 152 in 2008 to 138 in 2011 and 131 in 2017—a drop of about 14% in the county-level average number of loans in this size bin between 2008 and 2017. At the same time, instead, the average number of loans originated for all larger size bins has been increasing since 2011 throughout the rest of our sample period, until the end of 2017. We detect this increasing pattern for all larger loans regardless of whether they are above or below the baseline conforming loan limit of |${\$}$|417K, as well as above or below the exception high-cost-county conforming loan limits, which are not included in the range above |${\$}$|800K for virtually any U.S. counties.¹⁶ Based on the evidence in Figure 1, we conclude that the number of loans originated in the small and medium-sized groups has dropped over time, whereas the number of loan of larger sizes has increased over time.

Figure 1

This figure reports the change in lending by loan size bins for the years 2008-2017. Panel A reports the average number of approved loans, computed at the county level. Panels B and C reports the share of approved loans, computed at the lender level. In panel B the share is computed based on the overall number of approved loans. In panel C, it is computed based on the overall dollar value of approved loans. Within each panel, we report five plots for the following size bins: below |${\$}$|200K, |${\$}$|201K-|${\$}$|400K, |${\$}$|401K-|${\$}$|600K, |${\$}$|601K-|${\$}$|800K, and |${\$}$|800K and above.

Open in new tabDownload slide

The regressive redistribution of mortgage credit is also evident within lenders. In panel B of Figure 1, we compute, for each lender in our sample, the yearly shares of total loans they originate within each size bin. We then compute the weighted average of these shares across lenders based on the dollar value of lenders’ total origination. First of all, loans below |${\$}$|200K were the majority of loans originated in 2008. Loans in this bin have dropped substantially over time, moving from 57% to 41% of all loans between 2011 and 2017—a drop of 16 percentage points, or 28% of the initial share. The within-lender share of originations for all larger-size loans, instead, increases substantially. It increased from 31% to 43% for loans between |${\$}$|201K and |${\$}$|400K—an increase of 39% of the share in 2011. The increasing pattern is qualitatively similar in other size bins. For example, in the largest size bin (⁠|${\$}$|800K and above), originations increased from 0.8% in 2011 to 1.4% in 2017, which is a 75% increase relative to the share in 2011. Panel C of Figure 1 depicts the dynamics of within-lender mortgage origination across size bins based on the overall dollar value of originations instead of the number of loans. The patterns are qualitatively and quantitatively similar.

3.2 Fact 2: Top lenders have changed their origination behavior the most

Fact 1 shows that the origination of large loans has increased after 2011 relative to the period between 2008 and 2010. This behavior is unlikely to be similar across large and small lenders for a number of potential reasons. For example, smaller lenders are unlikely to be able to generate many large loans, given the size of their balance sheets and the freezing of the private-label securitization market after the financial crisis. On the other hand, large lenders, which operate in many business segments, such as proprietary trading and investment banking, may find it profitable to exit the mortgage origination business altogether and focus on the other segments. We therefore assess whether the post-crisis dynamics of mortgage origination differed based on lenders’ sizes.

Figure 2 plots the dynamics of origination separately across top-five and other lenders. Top-five lenders are defined as the largest lenders by origination each year. Panel A of Figure 2 shows that the pattern for the average number of loans across counties discussed above is amplified in counties that belong to the top-two quintiles based on the share of top-five lenders (“High,” solid green lines) relative to other counties (“Low,” dashed orange lines). For loans below |${\$}$|200K, counties in the High group drive the drop in originations from 2011 to 2017. As for the increase in lending in larger size bins, again counties with higher shares of top-five originators drove the increasing trend in approvals we documented in Figure 2.

Figure 2

This figure reports the change in lending by loan size bins for the years 2008-2017, separately by lender size. Panel A reports the average number of approved loans, computed at the county level, across counties with a share of top-five lenders in the top-two quintiles and other counties. Panels B and C reports the share of approved loans, computed at the lender level, across top-five lenders and other lenders. In panel B the share is computed based on the overall number of approved loans. In panel C, it is computed based on the overall dollar value of approved loans. Within each panel, we report five plots for the following size bins: below |${\$}$|200K, |${\$}$|201K-|${\$}$|400K, |${\$}$|401K-|${\$}$|600K, |${\$}$|601K-|${\$}$|800K, and |${\$}$|800K and above.

Open in new tabDownload slide

Panels B and C of Figure 2 show that the within-lender share of loans originated has decreased dramatically in the smaller size bin for all lenders. Whereas lenders below the top drive the increasing trend in origination after 2011 for the range |${\$}$|200K-|${\$}$|400K, which only includes loans below the baseline conforming loan limit, the trend is reversed for larger size bins. Indeed, the top-five lenders increased their origination of larger loans substantially more than other lenders. We confirm this pattern in panel C when considering origination in terms of dollar value of the originated loans. If anything, here the average dollar value top-five lenders provided for loans below |${\$}$|400K dropped. Instead, in the largest size bin, which does not include conforming loans in virtually any U.S. counties, top-five lenders increased their share of originations substantially after 2011, whereas smaller lenders did not.

This raw data time-series evidence suggest that all U.S. lenders moved to originating larger loans after 2011, but smaller lenders originated larger loans within the baseline conforming loan limit, whereas large lenders moved to originating larger loans without any constraints.

3.3 Fact 3: The change in origination has been proportional to lenders’ size

The third fact we document is a proportional relationship between lender size and change in mortgage origination behavior: the relationship does not change discontinuously at any lender size threshold.

Figure 3 presents the raw percentage change of loans originated by lender size between 2008 and 2017. The left panel considers loans below |${\$}$|417K. The right panel considers loans above |${\$}$|417K. We group institutions in 15 equal-size groups based on total lending, and report the value-weighted change in lending for each group. To avoid including very small banks, we limit the sample to the top 1,000 U.S. lenders. Every green dot in the plot thus comprises approximately 70 lenders. The larger is the lender, the stronger is the drop in the number of mortgages originated below the baseline conforming loan limit. At the same time, the larger is the lender, the higher is the increase in the number of loans originated above the baseline conforming loan limit. This relationship is not driven fully by one lender group or by a small number of lender groups. This fact suggests that regulatory or economic shocks that might have affected lenders based on specific size cutoffs cannot explain in full the dynamics we document.

Figure 3

This figure plots the raw percentage change in loans originated by lender size between 2008 and 2017. The left panel considers loans below |${\$}$|417K. The right panel considers loans above |${\$}$|417K. We group lenders in 15 equal size groups based on total lending, and report the value-weighted change in lending for each group. We limit the sample to the 1,000 top lenders.

Open in new tabDownload slide

3.4 Fact 4: Originations further from the conforming loan limit(s) drive the results

Lastly, we turn to assess whether the redistribution was homogeneous across loan sizes or if certain portions of the loan size distribution were affected more than others. In particular, we aim to understand if the incentives to bunch loan origination at the conforming loan limit (CLL) was the main driver of our findings (see DeFusco and Paciorek  2017).¹⁷

The baseline CLL was |${\$}$|417K from 2008 until the end of 2016, and was increased to |${\$}$|424K only in 2017, the last year in our analysis. Because the redistribution we document started in 2011 and has increased since, the baseline CLL change in 2017 cannot explain our results.

Note that, since the 2008-2009 financial crisis, exceptional CLLs were allowed to incentivize housing transactions in high-cost counties. The maximum high-cost-county CLL was |${\$}$|729.75K between 2008 and 2010, and |${\$}$|625.5K from 2011 to 2017. The maximum exception limit was thus higher in 2008-2010 relative to 2011-2017. The redistribution we document therefore cannot be explained by an increase of the high-cost-county CLL either, because the redistribution started after 2011, at a time in which exceptional CLLs were lower relative to 2008-2010.

To further assess how origination behavior has changed across the loan size distribution, in panel A of Figure 4, we plot the densities of the mortgage amounts originated by 3 of the top-five financial institutions throughout our sample period,¹⁸ based on the share of mortgages originated each year. In each graph, the long-dashed red line represents the density of loan amounts in 2008; the short-dashed lines represent the years 2011-2016; and the solid black line represents 2017. Changes in originating behavior are similar across large lenders. In 2008, all financial institutions had an incentive to originate loans below the conforming loan limit of |${\$}$|417K. Bunching of originated mortgages just below the value of the limit in 2008 emphasizes this incentive.¹⁹|$^{,}$|²⁰

Figure 4

Panel A reports the loan size distributions by three top originators by lending activity over the sample: Wells Fargo, J.P. Morgan, and U.S. Bank. Panel B reports the loan size distributions for all financial institutions that rank outside the top 20 (left plot), for the bank originators that rank outside the top 20 (middle plot), and for nonbank mortgage originators (right plot). Each plot reports densities for the years 2008, 2011, 2012, 2013, 2014, 2015, 2016, and 2017. Loan amounts are winsorized at the 0.5% level.

Open in new tabDownload slide

Since 2011, the distribution of mortgages shifted to the right for all lenders, and especially for larger lenders. Moreover, the incentives to bunch loans below the conforming loan limit has decreased dramatically for large lenders. The incentive was at its lowest in 2017, the last year in our sample, when the bunching was quite limited for Wells Fargo and U.S. Bank and completely inexistent for JP Morgan and Bank of America.

The drop in bunching is not evident for smaller lenders and nonbank originators (see the left plot of panel B of Figure 4). Even in this case, the distribution has started to shift to the right in 2011 and has not stopped since. But for smaller lenders, the part of the distribution below the baseline conforming loan limit has moved to the right. Instead, we barely detect any substantial changes in lending above the baseline conforming loan limit. We will analyze this pattern in more detail in the multivariate analysis below. This behavior is similar for both small banks and nonbank mortgage originators, as we report in the center and right plots of panel B of Figure 4.

To further assess whether our results hold true irrespective of the level of the conforming loan limit, Figure 5 limits the analysis to loans originated in high-cost counties. For this figure, we focus only on banks and mortgage originators in which mortgages satisfying a set of characteristics could be originated as conforming mortgages (and hence sold to GSEs) even for loan amounts above the basic conforming-loan limit of |${\$}$|417K during our sample period. Specifically, the exception counties we consider had an upper bound of the conforming loan limit of |${\$}$|729.75K for the years 2008-2010, which was then moved back to |${\$}$|625.5K starting in 2011.

Figure 5

The left plot shows the loan size distribution in counties with exceptions to the |${\$}$|417K conforming loan limit in 2010 and 2015 for loans originated by the Top 5 lenders based on U.S.-wide total origination. The right plot shows the loan size distribution in counties with exceptions to the |${\$}$|417K conforming loan limit in 2010 and 2015 for loans originated by all institutions outside the top-20 lenders based on U.S.-wide total origination. Loan amounts are winsorized at the 0.5% level.

Open in new tabDownload slide

In panel A of Figure 5, we compare the density of mortgage origination for the top-five lenders in the U.S. in 2010 and in 2015 in exception counties, which had a maximum conforming loan limit of |${\$}$|625.5K after 2011 and of |${\$}$|729.75K between 2008 and 2010.²¹ In panel B of Figure 5, we consider lenders outside the top five. In 2010 all lenders, irrespective of their size, had an incentive to bunch at the maximum exception counties’ conforming-loan limit of |${\$}$|729.75K, as we can see by the bunching of mortgage originations at such level. Interestingly, they also bunched at the baseline |${\$}$|417K conforming loan limit. In 2015, top originators moved without restrictions to originate larger loans, both below and above the |${\$}$|625.5K conforming loan limits. In fact, similar to our baseline results for all new originations in HMDA, the bunching at |${\$}$|625.5K was rather small.

Panel B of Figure 5 shows that lending for smaller lenders evolved differently. The bunching was very large and noticeable at the |${\$}$|729.75K maximum conforming loan limit in 2010, as it was at the |${\$}$|417K conforming loan limit. By 2015, the approved-loan density distribution had shifted to the right below the |${\$}$|417K, but there was still a pronounced bunching both at the |${\$}$|417K and the |${\$}$|625.5K limits. If anything, once the maximum conforming loan limit was moved down from |${\$}$|729.75K to |${\$}$|625.5K, smaller lenders dismissed the demand for loans they were fulfilling at the higher conforming loan limit in the 2008-2010 period and were constrained by the new maximum exception limit.

4. Mortgage Origination Since 2011: Multivariate Analysis

After describing the baseline facts in the raw data, we move on to present results that control directly for determinants of mortgage demand.

4.1 Extensive margin of mortgage origination

We first analyze the extensive margin of mortgage origination, that is, whether approval rates changed systematically depending on loan size since 2011 relative to the period 2008-2010. We also test whether this association was higher for larger lenders relative to smaller lenders and nonbank originators even after we keep constant observable determinants of mortgage demand.

This extensive margin analysis levers the most crucial advantage of HMDA data relative to other sources of data—the fact we observe the overall demand for loans, not only approved loans but all applications, including rejected applications. We can thus focus on the determinants of originators’ approval decision within the set of all applications. Our setting also allows abstracting from U.S.-wide and local time-varying shocks that might affect the aggregate demand and supply of mortgages by limiting our analysis within years and/or within county-year pairs.

We report the results from estimating Equation (1) in columns 1–3 of Table 2. In column 1, we limit the variation within counties to absorb any time-invariant county-level characteristics that might affect the demand and supply of mortgages locally, such as local regulation or industry distribution. We also limit the variation within years, to absorb U.S.-wide business cycle shocks, which are relevant during our sample period both before and after 2010. For the reference period 2008-2010, a 1% increase in the dollar value of the loan is associated with a 3.8-percentage-point higher likelihood of approval (⁠|$\hat{\beta}$|⁠), which is about 4.8% of the average likelihood of approval in our sample.

Over the period 2011-2017, the association between loan size and approval rate has increased by 55% (⁠|$\hat{\gamma}$|⁠). After 2010, a 1% higher dollar value was associated with a 6-percentage-point higher likelihood of being approved.

In columns 2 and 3 of Table 2, in which we account for time-varying local economic shocks and for additional determinants of mortgage demand, we still find that the association between loan size and likelihood of approval was 32% higher after 2010 relative to before.

Columns 4–6 of Table 2 report the results for this triple interaction specification. We see that indeed |$\hat{\delta}$| is positive: top-five lenders increased the likelihood of approval of larger loans more than other lenders after 2010. For smaller lenders, the likelihood of approval of loans that were 1% larger after 2010 was 2 percentage points higher relative to before 2010 (⁠|$\hat{\gamma}$| in column 4). For top-five lenders, the marginal increase in the likelihood of approval for larger loans after 2010 was 3.4 percentage points (⁠|$\hat{\gamma}$|+|$\hat{\delta}$| in column 4), or 70% larger than the association for smaller lenders.

Column 4 also shows that the probability of approval of loans by top-five lenders was 17 percentage points higher than other lenders up to 2010 (⁠|$\hat{\xi}$|⁠), and large lenders moved out of the mortgage space more than smaller lenders after 2010 (⁠|$\hat{\zeta}=-11.6\%$|⁠).

In columns 5 and 6 of Table 2, we verify that these associations are robust to absorbing time-varying county-level economic shocks as well as additional determinants of the demand for loans.

4.2 Robustness

Our multivariate analysis did not allow for time-varying effects of determinants of the demand for mortgages on mortgage origination behavior. For instance, demographic characteristics that were not relevant to the approval choice might have become more relevant since 2011. In column 1 of Table 3, we augment Equation (2) with a full set of interactions between our controls and the dummy |$After~2010$|⁠. The size and statistical significance of the triple interaction coefficient are, if anything, higher than the baseline ones (column 6 of Table 2).

We then test whether the time-varying concentration of lending in local markets might explain our results.²³ We follow Scharfstein and Sunderam (2016) and define local mortgage lending concentration as the market share of the top-four financial institutions by county-year. We find no significant correlation between our share of top-five U.S. lender activity and the measure of county-level lending market concentration (-0.001, |$p$|-value|$>$|85%). In column 2 of Table 3, we reestimate the triple interaction specification including the measure of local lending market concentration as well as its interaction with the dummy |$After~2010$|⁠. Consistent with the fact that the measures are barely correlated in the data, including these additional variables does not affect our results.

In column 3 of Table 3, we consider the possibility that time-varying shares of foreclosures explain our results, because large and small lenders might have different policies regarding foreclosures and these differences could also vary over time.²⁴ Unfortunately, we do not observe foreclosures at the local level in HMDA. We obtain this information from Zillow but only for a subset of U.S. ZIP codes, which reduces the sample size by about 40%.²⁵ Controlling for the share of local foreclosures does not change our results materially.

One might wonder whether banks and nonbank mortgage originators reacted differently to the change in incentives. For instance, Buchak et al. (2018) show FinTech mortgage originators proliferated after the financial crisis. In column 4 of Table 3, we exclude all loans originated by a nonbank lender, and confirm our baseline results.

Finally, we assess if the differential reaction to measures of unconventional monetary policy, rather than lenders’ size, explains our extensive margin results. DiMaggio, Kermani, and Palmer (2016) show that loans originated in sand states—the states more hit by the credit freeze before QE1—did not react to these measures. If we exclude all mortgages originated in sand states in column 5 of Table 3, we estimate a triple interaction coefficient of similar size and statistical significance as in the baseline analysis.

4.3 Intensive margin of mortgage origination

To interpret the results, consider the special case of the median, which is the 50|${th}$| percentile of the distribution. The coefficient |$\hat{\beta}(50)$| estimates that the median size of approved mortgages was |$\hat{\beta}(50)$| units higher since 2011 relative to 2008-2010. A positive |$\hat{\beta}(50)$| would suggest that the median of the distribution has shifted to the right. The advantage of estimating quantile regressions for the vast majority of the percentiles of the approved loan size distribution is that we can assess how the overall distribution of loans has changed over time, as opposed to only focusing on specific moments, such as the conditional mean.

Because estimating quantile regressions with samples as large as ours is computationally intensive, we report results for a random 50% of the sample. Using a subsample of the approved loans is immaterial for the coefficient estimates: we find virtually identical coefficients if using samples of 5%, 10%, or 25% of the observations. We report the results for estimating the set of regressions in Equation (3) in both graphical (Figure 6) and tabular (Table 4) forms. The graphical form allows easy comparison of the size of the coefficients across quantiles, whereas the tabular form reports the statistical inference associated with each estimate |$\hat{\beta}(\tau)$|⁠.

Figure 6

This figure plots the coefficients |$\hat{\beta}(\tau)$| estimated from a set of quantile regressions of the following form:

$$Q_{\tau}(Loan~size_{b.t}) = \alpha(\tau) + \beta(\tau) \ After~2010_{t}\times Top~5~lender_{b,t} + \gamma(\tau) \ Top~5~lender_{b,t} + \ X'_{i,b,t} \ \zeta(\tau) + \epsilon_{b,t},$$

where |$Q_{\tau}(Loan~size_{i.b.t}) $| is the |$q{\text{th}}$| quintile of the distribution of the size of approved loans by lender |$b$| in year |$t$|⁠; |$Top~5~lender_{b,t}$| is a dummy variable that equals one if lender |$b$| in year |$t$| is one of the top-five U.S. originators nationally, and zero otherwise; |$After~2010_{t}$| is a dummy variable that equals one in years 2011-2017, and zero in years 2008-2010; and |$X'_{i,b,t}$| is a set of average borrower characteristics at the lender and year level we observe in HMDA; |$\eta_{kt}$| is a full set of county-year fixed effects (note the year fixed effects absorb the level of the variable |$After~2010_{t}$|⁠). In panel A, we estimate these specifications only for the loans originated by the top-five lenders based on overall U.S. mortgage origination. In panel B, we restrict the sample to the loans originated by lenders outside the top-five group. In both of these panels, we plot the coefficient attached to the |$After~2010_{t}$| dummy. In panel C, we estimate these specifications for all loans in the sample, and plot the coefficient attached to the interaction |$After~2010_{t}\times Top~5~lender_{b,t}$|⁠.

Open in new tabDownload slide

Panel A of Figure 6 plots the estimated coefficients |$\hat{\beta}(\tau)$| for values of |$\tau$| between 5 and 99 for loans approved by top-five lenders. Two results emerge. First, every single quantile of the distribution, with the exception of quantiles 5 and 10, has increased since 2011. This fact suggests that the overall loan size distribution has shifted to the right for top-five lenders after 2010 relative to before. Second, the magnitude of the coefficients increases linearly with the quantiles of the distribution. The left panel of Table 4 shows that coefficient estimates grow with the quantiles and we can reject the null hypothesis that sizes are the same across most adjacent quantiles. Note also that coefficients increase linearly around the baseline conforming loan limit of |${\$}$|417K as well as at higher values. This result suggests that interpretations based on systematic differences across conforming and nonconforming loans are unlikely explanations for our findings. Otherwise, we should have observed substantial discontinuities around the conforming loan limits.

Panel B of Figure 6 plots the estimated coefficients for loans approved by smaller lenders and nonbank mortgage originators. Even in this case, all coefficients are positive, and hence the size distribution of approved loans has shifted to the right since 2011. Panel of Table 4 confirms that these size differences are statistically significant across most adjacent quantiles.

The main difference between the results for small and large lenders is that the size of the coefficient for small lenders is decreasing instead of increasing. This pattern suggests that smaller lenders increased the size of loans within the conforming loan limit, whereas larger lenders were more and more aggressive the larger the size of the loans. Because the average values are different across the two subsamples of large and small lenders, we cannot compare the sizes of the estimated coefficients directly across panels A and B of Figure 6. We propose additional tests for this comparison in the next section.

4.4 Intensive margin effect driven by loans far from the conforming loan limit

Panel C of Figure 6 reports the estimated coefficients |$\hat{\beta}(\tau)$| from Equation 4. We can interpret these coefficients as the differential change in the quantiles of the loan size distribution across lenders of the two size groups. Panel C shows that, for quantiles below the conforming loan limit, the estimated coefficients on the interaction are largely negative and their absolute value decreases the larger the quantile of the distribution. This pattern suggests that smaller lenders were more aggressive in shifting the distribution of originations to the right within the conforming loan limit relative to top-five lenders.

Instead, estimated coefficients switch to positive and increase in absolute value with the quantile of the distribution above the baseline conforming loan limit. Even this pattern seems consistent with the comparison between panels A and B of Figure 6: top-five lenders have increased the size distribution of their loans disproportionally more than smaller lenders in the nonconforming segment.

Panel C of Figure 6 also shows that the change in loan size distribution by lender types we documented in the raw data are driven by the two extreme parts of the loan size distribution—small- loans and very large loans. Instead, we do not observe large effects for loans just around the baseline or the exception conforming loan limits. We confirm that the differences across coefficients are not only economically significant but also statistically significant in the right panel of Table 4.

5. Instrumental Variable Strategy

The fact that large and small lenders are not assigned randomly to households hinders a causal interpretation of the facts we have discussed so far. Unobservables might have changed the distribution of lenders across counties since 2011 and might have also determined a change in mortgage demand since 2011. In this case, the change of the originating behavior of lenders would not be due to supply-side forces, but to demand. To disentangle the two effects, ideally we would match lenders and households randomly every year from 2008 to 2017, which is impossible.

A potential identification strategy should compare households with a high likelihood of matching with top-five lenders with other households that are similar in all respects, including unobservable characteristics. In particular, such households should have the same demand for mortgages, and should react similarly to changes in the supply of mortgages. The ideal source of exogenous variation in the likelihood of matching with large lenders should not be affected by the financial crisis and the developments of the mortgage market after the crisis.

To tackle this challenge, we propose an instrumental variable strategy. We instrument the yearly share of large lenders in a county—which captures the likelihood households demand mortgages from large lenders—in the period from 2008 to 2017 with the share of large lenders in the county as of 2007. The rationale is that the financial crisis or any developments in the mortgage market after the crisis could not have determined the share of large lenders in 2007.²⁶ Moreover, policy changes after the crisis could not have affected the share of large lenders in 2007. This instrument is likely to be relevant, because inertia is present in the spatial penetration of bank branches. We document the relevance of the instrument below.

Figures 7 and 8 describe graphically the variation in the share of top-five lenders in 2007. Figure 7 plots the probability density function for the share of large lenders. For the vast majority of counties, large lenders cover between 5% and 60% of the overall mortgage activity in 2007, and the modal value is about 23%. The variation in the share of large lenders across counties is substantial.

Figure 7

This figure reports kernel density estimates of the percentage of lending across counties originated by the top-five lenders based on the U.S.-wide yearly origination activity for the year 2007, which we use in our instrumental variable strategy.

Open in new tabDownload slide

Figure 8

This figure reports choropleth maps of the percentage of lending across counties originated by the top-five lenders based on the U.S.-wide yearly origination activity for the year 2007, which we use in our instrumental-variable strategy. Panel A focuses on the whole United States. Panel B focuses on the state of Iowa as an example of an otherwise homogeneous state displaying substantial variation in the share of top-five lenders across counties.

Open in new tabDownload slide

We find substantial variation in the share of top-five lenders even across similar bordering counties. Panel A of Figure 8 plots the spatial variation in the share of large lenders in 2007 across all U.S. counties. Panel B of Figure 8 plots the corresponding spatial variation for counties in Iowa. We zoom on Iowa, because Iowan counties are rather homogeneous in terms of observable determinants of the demand for credit, including the racial composition of local households, the median house prices, and average household income. Both panels document substantial spatial variation in the share of large lenders, including across areas that are otherwise similar.

5.1 Validity of the instrument: Exclusion restriction

If we want to interpret the results of our instrumental variable strategy causally, we need to assume an exclusion restriction. The share of large lenders in a county in 2007 should only affect the amounts lent in the county in the following years through the share of large lenders in the following years, and not through unobservable characteristics at the county level or individual-borrower level. This exclusion restriction cannot be tested directly, and hence we propose two sets of results to assess its plausibility.

First, we compare the trends of important determinants of the demand for mortgages across counties with different shares of top-five lenders in 2007. This test aims to diagnose our sample for potential differential pre-trends in the economic characteristics of counties treated with different shares of large lenders. It also allows us to test whether counties with higher or lower shares of large lenders behaved differently after the financial crisis, due to different resilience or to different distributional effects of the policy measures implemented after the financial crisis to support aggregate and local economic growth.

Figure 9 reports the results for this parallel trends test. In each panel of Figure 9, we split U.S. counties among those in the top-two quintiles of the distribution by share of top-five lenders (“High Top 5,” orange dashed lines) and those in the other quintiles (“Low Top 5,” green solid lines). In each panel, we allow for different scales of the |$y$|-axes when plotting the results for counties with a high share of top-five lenders (right |$y$|-axis) and a low share of top-five lenders (left |$y$|-axis). Of course, we keep the range of the two axes the same, so that using different axes merely shifts the time series above and below vertically without manipulating their shapes. Thus, we can directly assess whether the trends are parallel across these two groups of counties by visual inspection.

Figure 9

This figure plots the average of nine county-level economic variables from 2008 to 2017 across two groups of counties: counties in the top-two quintiles by the share of the top-five lenders in the county (High Top 5, right |$y$|-axis) and other counties (Low Top 5, left |$y$|-axis). In every plot, the two axes have the same range to ensure the trends can be compared directly. We describe the source of each variable in the main text.

Open in new tabDownload slide

The top row plots three alternative measures of income at the county level from three alternative data sources. First is the median household-level income in each county from HUD, which we average across counties in each of the two groups by share of originations by top-five lenders. Second is the average household income from the micro-data underlying the ACS, which includes in each year a representative 5% sample of the U.S. population. Third is the average per capita GDP at the county level from the BEA. Across these three alternative measures of income, we find striking parallel trends across the two groups of counties and over time, for both the period between 2008 and 2010 (pre-trend) and the period between 2011 and 2017 (post-trend). This fact suggests that counties with different shares of top-five lenders did not differ in terms of one of the most common indicator of local business cycles—local income per capita—neither during the financial crisis nor after the crisis.

The middle row of Figure 9 plots three indicators capturing variation in local labor markets, that is, the county-level unemployment rate from the BLS, county-level labor force participation from the ACS—which captures potential changes in the age composition of counties over time—and the average number of hours worked weekly by respondents to the ACS. Again, we detect striking parallel trends across the two groups of counties in such important indicators for the evolution of local business cycles over time.

Finally, the bottom row of Figure 9 plots three indicators related to housing across counties. First, we consider the share of homeowners in each county over time. Note that, contrary to all the other dimensions we have considered so far, our own results imply that this variable should have evolved differently across the two sets of counties. In fact, the share of homeowners should have evolved similarly across counties up to 2010. Starting in 2011, we should observe a drop in homeowners in all counties, but this drop should have been larger in counties with a higher share of top-five lenders. This is because all lenders dropped the amount of mortgage originations to households, but top-five lenders reduced their origination of conforming loans by more than smaller lenders and nonbank originators. Because conforming loans represent the vast majority of all the loans originated in the U.S., a drop in that margin should affect more individual potential homeowners than a drop in the higher part of the loan size distribution. Thus, when computing the average number of households that own their home, we should see a larger drop in this ratio for counties with a higher share of top-five lenders. Consistently, the left panel of the bottom row of Figure 9 shows the drop in the share of homeowners in 2011 was larger in counties at the top of the distribution by share of top-five lenders.

The last two panels of Figure 9 plot the average monthly rent nonhomeowners pay over time as well as the yearly growth rate of the value of local bank deposits over time. Again, we detect parallel trends across counties both before and after 2010.

In our second set of tests, we follow Chodorow-Reich (2014) and first compute the average of determinants of the demand for mortgages across groups of counties with higher and lower shares of top-five lenders in 2007. We then compare the differences in averages across groups to the variables’ standard deviations in the full sample. Small differences relative to the standard deviations suggest that the determinants are balanced.

Table A.1 reports the results for this test. The first four columns of Table A.1 report the average value of each variable for the observations in all years, split into four equal-size groups based on the share of top-five lenders in 2007. The first column refers to observations in the bottom quarter of share of large lenders by county, whereas the fourth column refers to observations in the top quarter of the share of large lenders by county. The fifth column reports the standard deviation for each of the listed variables. Variables include the average change in the mortgage amounts for loans in different size groups over the period 2007-2010. Moreover, we add dummy variables for whether the applicant is Black, Asian, or Latino; the share of Black, Asian, and Latino population in each county-year; the log of the median house price in each county-year; and the share of foreclosed houses in each county-year.

All variables appear to be balanced for varying shares of large lenders, as the differences in averages across groups of counties are substantially lower than the variables’ standard deviations in the full sample.

Although, as discussed above, the share of top-five lenders is not assigned randomly to counties over time, the results of our parallel trend analysis and balancing test seem to suggest that counties with different shares of top-five lenders behaved similarly throughout our sample period. This result contributes to attenuate the endogeneity concerns when implementing our instrumental variable strategy.

5.2 Two-stage least squares specifications

All the interaction terms that include the share of local originations by top-five lenders are instrumented with the corresponding interaction terms in which the dummy is replaced with the share of top-five lenders in the county as of 2007, before our analysis period. The county-level share of large lenders in 2007 is absorbed by the county fixed effect, because it does not vary within counties over time. For this reason, Equation (5) does not include the term |$\widehat{Share~Top~5~lender}$|⁠.

Because the specification includes two interactions and two excluded instruments, we estimated two (untabulated) first-stage regressions. We report two relevant first-stage statistics to assess the relevance and the weakness of our instruments at the bottom of Table 5. To assess whether our instrument is likely to be weak, we consider the Kleibergen-Paap (KP) Wald F-statistic, which is the version of the Cragg-Donald (CD) Wald F-statistic that allows for the adjustment of the clustering of standard errors at the county level in our baseline specifications. This statistic allows us to test the null hypothesis that our instrument is weak. For all IV specifications, we obtain Kleibergen-Paap Wald F-statistics of 45.195, 17.631, and 17.872, respectively. These values are substantially larger than those proposed by common rules of thumb, for example, an F-statistic above 10 is often suggested as indicative of unlikely weak instrument problems. Although these rules of thumb are not a definitive threshold for whether the issue of weak instrumentation is present or not, given the high values of our KP Wald-F statistics, we conclude that our IV procedure does not appear to face a weak instrument problem.

For a formal statistical test of the null of underidentified restrictions in the 2SLS procedure, we compute the Kleibergen-Paap (KP) statistics for the LM tests for underidentification. We have the following values for these statistics in our three specifications: 56.119 (⁠|$p$|-value|$<$|⁠.001); 20.132 (⁠|$p$|-value|$<$|⁠.001); and 20.306 (⁠|$p$|-value|$<$|⁠.001). We can reject the null of underidentification in all the specifications we propose at any standard level of statistical significance, and well below the 1% level in all cases.

Table 5 reports the second-stage estimates. We focus on |$\delta$|⁠, that is, the coefficient attached to the triple interaction between loan size, the dummy for the period after 2010, and the share of originations by top-five lender in the county-year. In column 1, the sign and magnitude of the estimated coefficient in the second state are similar to the baseline OLS results in column 4 of Table 2. The estimated IV coefficients are larger in magnitudes, but qualitatively similar to the baseline OLS results also for the specifications that add county-year fixed effect pairs. The increase in magnitude is due to the more stringent fixed effects specification that absorbs a larger portion of the variation of the instrument and in turn reduces the covariance between the instrument and the endogenous variable.

6. Assessing Potential Explanations

In this section, we discuss the extent to which a set of potential explanations are consistent with our findings. We consider two groups of potential channels: Risk-based channels and regulation-based channels.

6.1 Risk-based channels

6.1.1 Banks’ risk management

We start from considering the possibility that large lenders have moved towards jumbo loans more than smaller lenders to comply with capital requirements or to reduce the risk of their pool of assets. To assess this possibility, we first test directly whether the riskiest institutions have increased the origination of large loans and decreased the origination of conforming loans since 2011. We do so by recomputing Figure 3 using lenders’ riskiness instead of size to sort them. Our preferred measure of bank risk is the share of reserves over the total amount of nonperforming loans held by the bank. We report the results in Figure 10, which also replicates Figure 3 for the subsample of institutions for which we observe our risk measure. The pictures show no relation between riskiness and the change in lending behavior, whereas our baseline result is replicated for the subsample of banks for which we observe the risk measure.

Figure 10

This figure reports in panel A the raw percentage change in loans originated by lender risk between 2008 and 2017. The left panel considers loans below |${\$}$|417K. The right panel considers loans above |${\$}$|417K. The measure of lender risk is the share of reserves over the total amount of nonperforming loans held by the bank. Because we do not observe this measure of lender risk for all the institutions in our sample, we focus on the 100 riskiest institutions and group them into 10 equally sized groups based on lender risk. We report the value-weighted change in lending for each group. Panel B uses the same sample of lenders, but is sorted on size.

Open in new tabDownload slide

Note also that there is no evidence that jumbo loans are at lower risk of default than smaller loans, even if the applicants are on average wealthier. If anything, evidence based on customized CoreLogic data shows that, in the second quarter of 2010, the delinquency rate on mortgages for investment homes above |${\$}$|1 million was twice as high as the delinquency rate on mortgages below |${\$}$|1 million. Delinquency rates for large mortgages were higher than those for smaller mortgages for both owner-occupied and investment real estate (Streitfeld  2010).

6.1.2 Putback policy by GSEs

Second, we consider the fact that GSEs have introduced a stricter “putback” policy after the financial crisis, by which they started to force lenders to buy back loans whose origination documents were found to misrepresent borrowers’ ability to repay. This stricter policy created uncertainty in the effectiveness of the GSEs’ guarantee, and hence might have dis-incentivized lenders to originate conforming loans relative to jumbo loans.

To understand the extent to which putback risk might explain our facts we first need to clarify the timing and evolution of the GSE’s policies regarding mortgage repurchases: On March 29, 2010 (which is within our control period), Fannie Mae issued a new, stricter set of guidelines for the repurchase of loans by originators, which increased the quality standards for loans and hence increased the likelihood originators might face a repurchase request.²⁷ These stricter guidelines resulted in higher repurchase requests from Fannie and Freddie, and the complaints by mortgage originators that creditworthy applicants might be denied access to credit due to originators’ concerns about repurchase requests. In other words, this was the time when putback risk increased substantially for lenders based on this change in putback policy by GSEs. By contrast, on September 11, 2012, GSEs announced a first relaxation of the requisites for repurchase requests after substantial pressure on the side of mortgage originators. GSEs representatives as well as specialized industry publications, such as the American Banker, all agreed that the new guidelines would reduce the extent of repurchase risk of conventional loans for lenders (for instance, see: https://www.americanbanker.com/news/fhfa-plans-guidance-to-curb-putbacks). Finally, on November 20, 2014, GSEs announced new guidelines that relaxed the putback risk mortgage originators faced substantially,²⁸ as GSEs, lenders, and the press acknowledged.²⁹ These loosened guidelines were retroactive to include loans originated after January 1, 2013.

Overall, the timing of the major changes in GSEs’ putback policies suggests that this channel is an unlikely explanation for the facts we document. Indeed, putback risk increased in 2010, decreased a first time in 2012, and decreased substantially in 2014, with a retroactive effect back to 2013. The increase in risk thus happened within our control period, before we document any redistribution of mortgage credit origination. Moreover, putback risk decreased substantially within the period in which instead the redistribution has become starker and starker over time. Not only we do not observe any reduction in the redistribution phenomenon after putback risk decreased, which is what this economic channel predicts, but the redistribution continued to become even starker over time.

To further assess whether changes in putback policies over time could explain our results, we consider the variation in the amount of mortgage repurchases GSEs imposed to lenders over our sample period. For this explanation to be relevant in our case, we would need to observe that the repurchase requests have increased or at least stayed flat over time. This is because a decreasing trend in repurchase requests would signal lenders that the policy is less stringent over time and by construction cannot explain the redistribution we document, which instead becomes larger and larger over time. In fact, this is only true to the extent that the quality of the pool of borrowers has not changed systematically across large and small lenders in the years after 2011. This possibility is consistent with our results reported in Section 5, which document a similar evolution of several determinants of the demand for mortgages across counties with a different presence of top lenders.

In Figure A.3 in the Internet Appendix, we plot the overall dollar amount of the mortgage repurchase requests Freddie Mac imposed from 2008 to 2015.³⁰ The graph shows two salient facts for our purposes. First, the peak of the repurchase requests was in 2010, which is part of our control period. Second, the repurchase requests have been trending down since 2010 and, on average, they have been substantially lower in the years after 2010 relative to the years before 2011.

Because repurchase requests have been trending down and decreased substantially over time, from about |${\$}$|4B in 2010 to less than |${\$}$|0.5B in 2015, they cannot explain why lenders, and especially large lenders, have kept redistributing mortgage origination from smaller loans to very large loans more and more since 2011.

6.1.3 Private label securitization market

During the 2008-2009 financial crisis, private-label securitization came to a halt (Goodman  2015). As documented in Goodman (2015, figure 2), the aggregate amount of issuance of private-label residential mortgage-backed securities has dropped from approximately |${\$}$|700 billion in 2007 to approximately |${\$}$|60 billion in 2008. Private-label residential mortgage-backed securities have remained stagnant throughout the period 2008-2017, even if private-label securitization in other asset classes, such as credit cards, automobile loans, and student loans, started to recover over time. The fact that private-label residential mortgage securitization did not recover throughout the period 2008-2017 is an indication that this channel cannot explain our results.

6.1.4 Change in the risk profile of small-dollar loans

The default risk of small-dollar loans—loans of about |${\$}$|70K or lower—was higher than the default risk of loans between |${\$}$|70K and |${\$}$|150K before the 2008-2009 financial crisis.³¹ However, the two risk profiles have converged after the crisis, to the extent that they are indistinguishable since 2012. Note that, if this explanation was relevant, it would still only help us understand the redistribution we document partially, because it would only apply to the loan-size distribution for small and medium-sized loans.

In a partial equilibrium view, if lenders originated loans based on right-on-time risk assessment models, which is what all lenders argue they do, the origination of small-dollar loans should have, if anything, increased relative to loans between |${\$}$|70K and |${\$}$|150K since 2011, which is the opposite of what we document in this paper. At the same time, the very fact that lenders imposed stricter standards to the origination of small-dollar loans after the crisis might explain the convergence in default risks. In this case, we should observe a discontinuity in redistribution between small-dollar loans (⁠|${\$}$|70K or lower) and slightly larger conforming loans (⁠|${\$}$|70K and |${\$}$|150K), which though we do not detect for either large or small lenders.

6.2 Regulation-based channels

From the 2008-2009 financial crisis onwards, several policies and changes in financial regulation were implemented, some of which might have affected lenders’ incentives to originate mortgages.

6.2.1 Differential effects of unconventional monetary policy across lenders

On the monetary policy side, Chakraborty, Goldstein, and MacKinlay (2020) document the effects of mortgage-backed securities (MBS) repurchase waves by the Federal Reserve (Fed) on banks’ lending behavior. They find that banks increase lending as a result of this policy, and the extent of the reaction differed across lenders. If the change in lending following MBS repurchase waves also coincided with a redistribution of lending across the size distribution, especially for large lenders, this channel would help explain our results.

We propose a test for assessing this explanation by considering the fact that MBS repurchases by the Fed happened in waves throughout our sample period. Chakraborty, Goldstein, and MacKinlay (2020) show that the policy was most effective in terms of increasing lending in years that followed fourth-quarter MBS purchases relative to other years. We thus estimate our baseline triple-interaction specifications from Equation (2) separately for loans originated in years in which the policy was most effective or in other years. If reaction to MBS repurchases drives our results, we would expect that the size of the estimated coefficient on the triple interaction of loan size, the dummy for the period after 2010 relative to before 2010, and the dummy for top-five lenders is significantly larger in the years in which the policy was most effective relative to other years.

Table 6 reports the results for this test. We see that across all three baseline specifications, if anything, the estimated coefficients are larger in years in which the MBS repurchase policy was least effective, which seems inconsistent with a substantial role of this policy in explaining the redistribution we document. Note that we instead we replicate the increase in the overall amount of lending in effective years, which is totally consistent with Chakraborty, Goldstein, and MacKinlay (2020). Reaction to this form of unconventional monetary policy though does not seem to explain the regressive redistribution from smaller to larger loans within lenders.