RDP 2013-05: Liquidity Shocks and the US Housing Credit Crisis of 2007–2008 7. Robustness Tests

The key hypothesis of the paper is that the OTD lenders became liquidity constrained when the securitisation market effectively shut down in 2007, and this increase in liquidity constraints, in turn, caused the OTD lenders to reduce new mortgage lending disproportionately. A key assumption underpinning this hypothesis, and the identification strategy, is that the OTD and non-OTD lenders are similar in all respects except the extent to which they became liquidity constrained during the crisis. There remains some possibility that the characteristics of OTD and non-OTD lenders differ along other unobservable dimensions, and that these differences could be driving the variation in lending behaviour across the two lender groups during the crisis. For example, OTD lenders may have been more willing to take risk or they may have been more reliant on specific loan origination channels that may have been associated with greater risk-taking (e.g. mortgage brokers). It could be these systematic differences, rather than differences in financing constraints, causing the observable variation in lending behaviour during the crisis. The purpose of the following series of tests is to rule out such alternative explanations for the observed link between securitisation and mortgage lending during the crisis.

7.1 Changes in Mortgage Lending Standards

Some empirical studies suggest that securitisation contributed to bad lending by reducing the incentives of lenders to carefully screen borrowers (Mian and Sufi 2009; Keys et al 2010; Rosen 2010; Purnanandam 2011). These studies argue that securitisation weakened lenders incentives to screen borrowers by making the link between loan originators and the investors who bear the default risk more opaque. It also led to asymmetric information between loan originators and final investors and, subsequently, moral hazard problems. This suggests that OTD lenders may have had weaker incentives to screen borrowers than non-OTD lenders. It could be that this pre-crisis difference in lending standards between the two groups, and not differences in financing constraints, caused difference in lending behaviour during the crisis.

I need to make some subtle changes to the estimating equation to test the robustness of my results to this alternative lending standards explanation. To start, I re-write the equation in levels:

There are a couple of small, but important, differences between Equations (1) and (4). First, the subscript on the loan sale variable (SALESHARE_ij) indicates that the reliance on loan sales varies by lender and tract, rather than just lender. The highly disaggregated nature of the loan-level data means that the share of loans that are sold can be constructed for each lender in each tract. Second, the additional variation in the loan sales share variable due to this re-specification means that lender-specific time trends (η_it) can be separately identified within the error term. The lender-specific time trends control for unobservable bank-specific factors that vary systematically over time, such as changes in bank lending standards. The original specification could not include lender-specific dummies as these would be perfectly collinear with the sales share variable, which only varied by lender. This equation can again be written in growth rates by taking first differences over time:

As discussed earlier, the OLS estimator of ρ is biased if unobservable credit demand shocks are correlated with a lender-tract's reliance on loan sales (i.e. corr(SALESHARE_ij,Δη_j) ≠ 0). But, as before, this is easily handled by including tract dummies in Equation (5). However, the OLS estimator of ρ can now also be biased if unobservable credit supply shocks are correlated with a lender-tract's reliance on loan sales (i.e. corr(SALESHARE_ij,Δη_i) ≠ 0). For example, if banks that were reliant on the originate-to-distribute model were also more likely to reassess their risk exposures and tighten lending standards during the crisis there could be a negative correlation between a lender-tract's reliance on loan sales and bank-specific lending growth (i.e. corr(SALESHARE_ij,Δη_i) < 0). By including a dummy variable for each lender in the growth rate equation, I control for unobservable changes in lending policies across financial institutions, including changes in lending standards. In other words, I can identify the relationship between the pre-crisis reliance on securitisation and mortgage lending during the crisis after controlling for both unobservable tract-specific and lender-specific shocks. This eliminates any potential source of endogeneity caused by differences in the national lending policies of OTD and non-OTD lenders. The results of estimating Equation (5) are shown in Table 4.

The results of estimating the model using the lender-tract share of loan sales are very similar to the benchmark model that uses the lender share of loan sales. Based on the specification with just tract fixed-effects (column 2) the coefficient estimate is −0.108, which is similar to the estimate of −0.077 obtained from the benchmark fixed-effects model (column 2 of Table 2). More importantly, the negative relationship between pre-crisis securitisation and lending activity during the crisis still holds even when I include both tract and lender fixed effects (column 3). In other words, the OTD lenders do not appear to have disproportionately reduced lending because of a relatively large (unobservable) tightening of bank lending standards. Rather, the link between loan sales and lending activity remains even after controlling for changes in bank lending policies. The estimates from the specification with both lender and tract fixed effects (column 3) suggests that a one standard deviation shock to the share of loans sold is associated with a 21 per cent decline in new mortgage credit.

7.2 Private versus Public Securitisation

A key assumption underpinning the benchmark model is that US mortgage lenders cannot easily substitute between different sources of funding, so the lenders that were dependent on securitisation would have become more liquidity constrained when the private-label market shut down in 2007.^[14] But this assumption may not hold if the mortgage lenders were able to substitute towards other less-affected sources of finance. For example, lenders that were particularly reliant on privately securitising mortgages may have sold their loans to the GSEs instead. The flow of funds data presented earlier suggested that, in aggregate, there was a substitution away from private securitised lending to public securitised lending during the crisis as the GSEs stepped into the breach caused by the disruption to the private-label market. In other words, the flow of funds evidence suggests that at least some lenders were able to substitute away from the worst-affected sources of finance.

The HMDA information can be used to identify each lender's reliance on both public and private loan sales. If the liquidity constraints hypothesis is true, the lenders most reliant on private securitisation in the pre-crisis period should have become more liquidity constrained than lenders dependent on public securitisation and hence would have scaled back credit by relatively more during the crisis.

To examine this, I re-estimate the benchmark equation but, for each bank, I split the share of loan sales into two components – the share of loans that are sold to the GSEs (PUBSHARE) and the share of loans that are sold to private financial institutions (PRIVSHARE):

If the liquidity constraints hypothesis is true, then there would be a significant negative effect of the private sale share variable (PRIVSHARE) on lending (i.e. β₁ < 0). Moreover, the effect of private securitisation would be greater than the effect of public securitisation (β₁ < β₂). If, instead, public securitisation had a larger (negative) effect on lending, then this may be evidence of a confounding factor, related to the OTD business model, causing all such lenders to cut credit. The results are shown in Table 5.

A comparison of the coefficient estimates on the private and public sale share variables (rows 1 and 2) suggests that the effect of the liquidity shock on lending activity is significantly larger for the private sale share than for the public sale share. This is confirmed in a separate t-test that directly compares the two coefficient estimates. The estimated effect of the private sale share on lending activity is also economically larger than the effect for the total sale share shown in Table 2. A one standard deviation shock to the private sale share variable is associated with a 16 per cent decline in new mortgage credit. This supports the hypothesis that OTD lenders reduced lending because they became financially constrained in 2007 and not because of some other unobservable confounding factor affecting all lenders that are reliant on the originate-to-distribute model.

7.3 Afilliated Non-bank Mortgage Lenders

To test the relative merits of the liquidity constraints hypothesis, I construct another test in which I focus specifically on OTD lenders. Many OTD lenders are mortgage companies that specialise in home mortgage lending, whereas most non-OTD lenders are depository institutions. These mortgage companies typically rely solely on securitisation to finance their lending, and generally cannot fund themselves through alternative sources of finance, such as deposits. However, there are important differences within the pool of OTD lenders. For instance, some mortgage companies are affiliated with depository institutions, such as Citigroup, whereas others are not. The mortgage companies that are affiliated with a bank potentially have access to a more diversified funding base (through internal capital markets) than mortgage companies that are not affiliated. We might therefore expect the affiliated mortgage companies to be relatively less vulnerable to a funding shock that is specific to the securitisation market than the non-affiliated companies.

This variation across OTD lenders in the ability to diversify funding risk allows me to test the liquidity constraints hypothesis against alternative explanations for the link between pre-crisis reliance on securitisation and post-crisis mortgage lending. If the liquidity constraints hypothesis is true, the lending of non-affiliated mortgage companies (that are solely reliant on securitisation) should be more responsive to the liquidity shock than the lending of affiliated mortgage companies. So I re-estimate the benchmark equation to test the liquidity constraints hypothesis, but now focus on the subset of mortgage companies (that are predominantly OTD lenders):

where I include a dummy variable for whether the lender is affiliated with a commercial bank or not. The dummy variable NONAFFILIATED takes a value of one if the lender is not affiliated with a bank and is zero otherwise. The dummy variable is interacted with the share of loans sold by each lender. All the other variables are as before. Under the liquidity constraints hypothesis, the negative effect of the liquidity shock should be larger for the non-affiliated lenders (i.e. β₁ < 0). The results are shown in Table 6.

In both columns 1 and 2, the coefficient estimate on the interaction term NONAFF * SALESHARE is negatively signed. This suggests that the non-affiliated mortgage companies, which lacked alternative funding sources, had a greater tendency than the affiliated mortgage companies to reduce lending in response to the liquidity shock. However, the coefficient estimate on the interaction term is not statistically significant. So, overall, the results only provide tentative evidence to support the liquidity constraints hypothesis.

7.4 The Aggregate Effect of the Liquidity Shock

The ‘within-tract’ identification strategy does not provide the complete picture of the aggregate effect of the liquidity shock on mortgage lending. This is because the strategy implicitly does not allow borrowers to substitute between different lenders (where Census tracts are thought of as ‘borrowers’). But borrowers might compensate for any reduction in credit from OTD lenders by obtaining alternative finance from non-OTD lenders. This substitution towards unaffected lenders could limit the effect of the liquidity shock on aggregate mortgage lending.

One approach to identify the aggregate effect of the liquidity shock on mortgage lending would be to estimate the relationship between the tract-level lending growth and the tract-level share of loans sold, which would implicitly allow borrowers to substitute between lenders. As discussed earlier, the estimates from such a regression will be biased if changes in total mortgage credit at the tract level reflect both changes in credit demand and supply.

Jimenez et al (2011) have recently proposed a method to adjust these estimates for the bias using the (unbiased) coefficient estimates obtained at the lender-tract level. This approach effectively separates the impact of supply from demand while allowing borrowers to substitute between lenders. The approach is described in more detail in Appendix C. To do this, I estimate the tract-level version of Equation (2):

where denotes the log change in credit for tract j across all mortgage lenders. It is essentially a weighted average of the growth rate of credit at the lender-tract level, where the weights are given by each lender's share of loans within each tract. Similarly, denotes the (weighted) average pre-crisis reliance on loan sales of lenders that grant credit to tract j. The specification includes a set of tract-level control variables, such as the average income growth of loan applicants (). The same credit demand shock (Δη_j) appears in Equations (2) and (8) assuming that the shock affects a tract's borrowing from each lender equally. I then adjust these estimates using the following formula (which is outlined in Appendix C):

The results of estimating Equation (8) are shown in Table 7. As the first row of the table indicates, the OLS estimate of the aggregate effect is −0.364. Recall that the lender-tract level OLS estimate is −0.094 and the fixed-effects estimate is −0.077. Moreover, the data suggest that the sample variance of the share of loans sold at the lender level (V(SALESHARE_i)) is about 0.012 while the sample variance of the share of loans sold at the tract level (V()) is about 0.116. Combining all these estimates, and using the adjustment formula, the unbiased estimate of the aggregate effect of the liquidity shock is −0.207 (that is, −0.207 = −0.364−(−0.094+0.077) * (0.012/0.116)).

Overall, the results imply that a one standard deviation increase in the share of loans sold in aggregate is associated with a decline in new mortgage lending of around 2.3 per cent, on average. In aggregate, new mortgage lending fell by around 16.7 per cent, so a one standard deviation increase in the share of loans sold can explain about 14 per cent of the aggregate fall in new mortgage credit. This estimated (general equilibrium) effect is very similar to the (partial equilibrium) effect identified at the more disaggregated lender-tract level. This suggests that there was very little substitution between OTD and non-OTD lenders by borrowers following the liquidity shock.

Footnote

There is an additional assumption that borrowers are unable to perfectly offset funding shocks by substituting towards other sources of external finance. This assumption is more likely to hold in housing finance than in corporate finance because corporations typically have greater access than households to other funding sources (e.g. public debt and equity markets). Moreover, it is generally costly to re-apply for credit if the borrower's initial application is rejected. In other words, if there is a supply-side effect of the liquidity shocks, it is likely to be particularly important for household lending. [14]