P2P Loans and bank loans, the chicken and the egg, what causes what?: further evidence from a bootstrap panel granger causality test

ABSTRACT

This study attempts to re-investigate the causal link between bank loans and Peer-to-Peer (P2P) loans from China using data sets from eight areas (i.e., Beijing, Shanghai, Jiangsu, Zhejiang, Shandong, Hubei, Guangdong and Sichuan) over 2014M1-2016M4. We apply a bootstrap panel causality analysis that considers both cross-dependency and heterogeneity across cities. The empirical results indicate a unidirectional Granger causality running from P2P loans to bank loans for Beijing, Shanghai, Zhejiang and Shandong; feedback between P2P loans and bank loads for Jiangsu only and independence for the other three areas (i.e. Hubei, Guangdong and Sichuan).

KEYWORDS:

• P2P Loan
• bank loan
• bootstrap panel Granger causality test
• slope and homogeneous tests

JEL CLASSIFICATION:

• C33
• E44
• E52

Disclosure statement

No potential conflict of interest was reported by the authors.

Notes

¹ Since area-specific bootstrap critical values are used, the variables in the system do not need to be stationary, implying that the variables are used in level form regardless of their unit root and cointegration properties. Therefore, the bootstrap panel causality approach does not require any pre-testing for panel unit root and cointegration analyses. In addition, by imposing area-specific restrictions, we can identify which and how many areas exist in the Granger causal relation (see Kónya 2006).

² It is important to note here that, since the results from the causality test may be sensitive to the lag structure, determining the optimal lag length(s) is crucial for the robustness of our findings. As indicated by Kónya (2006), the selection of the optimal lag structure is of the utmost importance because the causality test results may critically depend on the lag structure. In general, both too few and too many lags may cause problems. Too few lags mean that some important variables are omitted from the model, and this specification error will typically cause bias in the retained regression coefficients, leading to incorrect conclusions. Conversely, too many lags waste observations, and this specification error will typically increase the SEs of the estimated coefficients, making the results less precise. For a relatively large panel, equations and variables with a varying lag structure would lead to a substantial increase in the computational burden. In determining the lag structure, we follow Kónya’s approach in which maximal lags are allowed to differ across variables but be the same across equations. We estimate the system for each possible pair of , , and by assuming from 1 to 4 lags, respectively, and then choosing the combinations that minimize the Schwarz Bayesian Criterion.

³ For the bootstrap procedure on how the region-specific critical values are generated, interested readers can refer to Kónya (2006).