RDP 2015-06: Credit Losses at Australian Banks: 1980–2013 Appendix C: More Regressions

C.1 Robustness

The results of Model C are robust to a number of alternative specifications. In particular:

• The choice to interact business sector variables with the business share of banks' portfolios is supported statistically. If interactions between the household share of the portfolio and these variables are added to a parsimonious version of Model C (e.g. (PSL_{i,t − 1} + HSL_{i,t − 1}) × business sector interest burden), they are not significant. Some lending to small businesses in Australia is collateralised by residential property, but an interaction between the business share of lending and residential property price growth is not statistically significant if added to Model C.
• For variables other than credit and loan growth, lagged relationships longer than one year are not a common finding in the literature (see, for example, Hess et al (2009)), and altering the lag structure of Model C changes little. A model with both contemporaneous and lagged values of business profits growth, household disposable income growth, and property prices, leads to very similar estimated coefficients and significance. The only exception to this is that the first lag of household income growth becomes significant at the 5 per cent level. But this variable is only marginally economic significant – a one standard deviation fall in household income raises credit losses by 13 basis points for a representative bank, versus 24 basis points for a one standard deviation fall in business profits.
• Ordinary least squares and random effects models yield coefficient estimates and standard errors very similar to the fixed effects version of Model C in Table 2. The only exception is the importance of the interaction between the personal share of lending and the household sector interest burden. These models estimate a strong and significant (at the 5 per cent level) relationship between the household sector interest burden and losses on personal lending. This result is intuitive, but is not particularly important at the bank level, given the low portfolio share of this type of lending. A one standard deviation increase in this macro-level variable increases credit losses by roughly 5 basis points for a representative bank (with 10 per cent personal lending).
• Economy-wide variables such as GDP growth, changes in the unemployment rate and inflation are insignificant if added to Model C, even if interacted with portfolio shares.
• Several key macro-level variables remain significant if the estimation sample is restricted to a period that excludes the early 1990s downturn. For example, in a model estimated using data for 1997–2013, the business sector interest burden and commercial property prices remain statistically significant and have coefficients similar to Model C, but business profits growth and bank-level loan growth become insignificant.
• A dynamic model including a single lag of current loss ratio was estimated using the Arellano and Bond (1991) estimator. The lagged dependent variable has a positive estimated coefficient and was significant at the 10 per cent level, though unlike for stocks of non-performing assets, there is no strong reason to expect true state dependence over short horizons in the flow of current losses. Estimated coefficients on key explanatory variables in this model were quantitatively similar to Model C. Banks learning from their mistakes may create a negative relationship between credit losses in an earlier downturn and those in a later downturn – something akin to the ‘institutional memory hypothesis’ of Berger and Udell (2004). But only a small number of the banks in the sample during the global financial crisis episode were not present during the early 1990s downturn, so testing this hypothesis is difficult.
• Clustering standard errors two ways – by bank and by year – leads to no substantive change in results. This approach is likely not entirely robust, given the number of clusters in both dimensions is below 50 (Cameron and Miller 2013).

C.2 Additional Regression Outputs