RDP 2021-01: The Role of Collateral in Borrowing 3. Empirical Strategy

Our objective is to compare how collateralised and uncollateralised markets respond to system-wide stress, and how the responses depend on borrowers' and lenders' counterparty risk and collateral availability. To achieve this we regress interbank activity at the lender-borrower-day-market ( lbdm ) level in each market, and the difference in activity across markets, on an interaction between market-wide stress and ex ante bank characteristics. The idea is that following unexpected system-wide stress, differences in risk characteristics and available collateral levels lead to unanticipated differences in access to credit in the two markets. The within-market aspect of this approach is similar to studies of heterogeneous effects of monetary policy changes (e.g., Kashyap and Stein 2000; Jiménez et al 2014) and to analysis of US interbank markets around the Lehman Brothers default (Afonso et al 2011). With the cross-market dimension, the strategy can be interpreted as a triple differences-in-differences with heterogeneous treatment effects.

Identification rests on an assumption that the measure of market stress is exogenous, that is, it is not driven by the bank behaviour that we analyse. We have high confidence that conditions in Australian interbank markets had no material contribution to the Lehman Brothers default, owing to the difference in size and global centrality between the US and Australian financial systems. This is an advantage of focusing on the Australian interbank market. To maintain exogeneity of borrower (and lender) counterparty risk and collateral holdings, they are measured at the start of the sample or earlier. By focusing the sample on a relatively tight window around the Lehman Brothers default, we ensure that the pre-sample measure of collateral holdings – a characteristic that can endogenously change relatively quickly – is both relevant and exogenous.

We first examine three ‘individual’ markets: repo, unsecured, and treating both as a single market (i.e. summing lbdm -level observations across markets to produce lbd -level variables). The explanatory variables of interest are the interactions between the system-wide stress measure and the counterparty characteristics represented by Xi , where

X i =[ CPRis k i Collatera l i CPRis k i *Collatera l i ]

To ensure that estimated coefficients are not biased by endogenous counterparty selection, we include counterparty*day fixed effects (e.g. when analysing borrower characteristics, lender*day).

The theoretical literature argues that monitoring incentives and diversification motives are important for interbank transactions (Rochet and Tirole 1996; Freixas and Rochet 2008; Allen and Gale 2009), which implies that lenders choose their borrowers in similar (for monitoring) or in different (for risk diversification) businesses and geographical areas. These fixed effects ensure that, for example, if a lender's behaviour reacts to system-wide stress, it is not attributed to the borrowers that it lends to. When analysing borrower characteristics, we also include borrower fixed effects (and similar for lenders), so that all time-invariant bank-level unobservables are controlled for. Our individual-markets regression specification for analysing borrower characteristics is

(1) LoanValue s lbd = α ld + α b +L B d * X b β+ ε lbd

To analyse lender characteristics, the l and b subscripts are swapped.

Next we formally compare the difference between the repo and unsecured markets. The variables of interest are triple interactions between an indicator variable for whether the market is collateralised ( 1s ), market stress ( LBd ), and counterparty characteristics ( Xi ). Following the credit channel literature, we saturate the regressions with fixed effects to control for unobserved variation (e.g. Khwaja and Mian 2008; Jiménez et al 2017); in our case lender*borrower*day fixed effects. This isolates the difference between the two markets, removing common variance driven by, for example, time-varying counterparty relationships. The coefficients on these triple interactions tell us whether, in reaction to market stress, the allocation of borrowing between repo and unsecured markets changes more for certain types of borrowers than others. In other words, whether substitution between markets is determined by bank-level characteristics. To analyse borrower characteristics, the regression specification is

(2) LoanValue s lbdm = α lbd + 1 s ( γ+ X b θ+ϕL B d +L B d * X b β )+ ε lbdm

where β is the coefficient vector of interest. Again, to analyse lender characteristics, the l and b subscripts are swapped.

The regressions are also repeated with the dependent variable replaced by Participationlbdm using a linear probability model. This specification reveals whether changes in LoanValueslbdm also took place at the extensive margin, that is, whether borrowers (lenders) changed the number of lenders (borrowers) they dealt with. Specifically, if a borrower borrows from N different lenders throughout the sample (summarised in Panel C of Table 1), an estimated coefficient of x implies a change in the number of lenders on a given day by xN.[7]

As a placebo test, the regressions are re-run on the corresponding 2006 sample, using 2008 values for the Lehman Brothers default treatment variable, in line with the suggestion of Roberts and Whited (2013). All loan-related and bank-characteristic variables are constructed as described in Section 2.4, but using 2006 data. The sample also starts on the second Monday in September, running for four weeks until Friday 6 October (with 19 days owing to a public holiday in this period).

This also demonstrates robustness against seasonal factors such as quarter-end or time-of-month effects.


That is, there are N observations for that borrower in that market each day, and x is the change in probability of each observation being one rather than zero. [7]