RDP 2021-01: The Role of Collateral in Borrowing 5. Further Robustness

This section reports the results from several additional robustness tests. We first run a set of placebo regressions. These repeat the main parts of the analysis in Section 4 on a 2006 sample, when markets were calm, covering four weeks from the second Monday in September (see Figure 3 for the market aggregates). The placebo regressions produce null results (Table B2). This addresses two potential concerns. First, it shows our results are not driven by patterns in interbank markets that occur on a regular basis, such as time of month, time of quarter, or time of year effects. Second, it shows that our results are not driven by random variability.

To test whether counterparty risk is exogenous to omitted variables, we regress borrowers' NPL on their size, leverage, domicile (an indicator equal to one if the bank is not Australian) and quantity of interbank borrowing in the week prior to our sample. The coefficient on pre-sample borrowing tests whether counterparty risk is endogenous to interbank activity prior to the peak-stress period. Only domicile is statistically significant (Table B3) – its p-value is 0.075, while the next lowest p-value is 0.324, on leverage. To test whether the counterparty risk coefficients in our main regressions are instead picking up domicile effects, we repeat the regressions after replacing counterparty risk with the domicile dummy variable. The borrower results are largely null (Table B4), which further supports our use of NPL as a measure of counterparty risk.

Next, to test the robustness of LBd as a measure of stress, we repeat the main regressions after replacing LBd with the TED spread, lagged one day to account for the time zone difference. The results have similar signs and significance levels (Table B5). Finally, we re-run the main regressions using different standard error clustering levels. The regressions in Section 4 cluster at the borrower and lender and day levels, which tends to make the minimum number of clusters around 15 to 20. To ensure that these are not presenting unconverged estimates, we re-run the regressions such that none have fewer than 30 clusters. For the specification (1) regressions, we cluster at the borrower*lender and borrower*day and lender*day levels simultaneously, and for the specification (2) regressions, we cluster at the lender and borrower levels. The results have similar significance levels (Table B6).