Evaluating discrete choice model specifications in SAFE-based research: implications for research in SMEs access to bank finance

Small and medium-sized enterprises (SMEs) access to bank finance is a significant concern for researchers, with much literature using discrete choice models and the ECB/EC Survey on the Access to Finance of Enterprises (SAFE). However, there is no consensus in this literature on the treatment of standard errors. Some employ heteroskedastic robust standard errors, some cluster standard errors at the country level, while others cluster at the country x time level. Yet, different standard error treatments can lead to different effects and can affect the reliability of model estimations. The methodology employed is a discrete choice binary dependent probit model and sensitivity analysis, which subjects the main findings to different standard error treatments and the application of a novel diagnostic framework to choose the best performing discrete choice model. The main findings show that results for variables from SAFE remain consistent regardless of standard errors used, but country effects vary with different treatments. This highlights the need for researchers to make explicit their research choices and rationale regarding standard errors, subject their findings to sensitivity analysis and ensure valid inference. In general, clustering at the country x time level is particularly appropriate when using discrete choice models when the data exhibits both cross-sectional and temporal dependencies. This paper contributes to the literature by examining standard error treatments used in previous SAFE studies and introducing a diagnostic framework to identify the best-performing discrete choice model. This diagnostic framework bridges the gap between econometric guidance and applied studies using SAFE and can be applied more generally to multi-country discrete choice analysis.