Is it possible to detect Tax Evasion using administrative data? A proposal based on a Binary Choice Model Affected by a Severe Censoring Mechanism
Tax evasion is a matter of huge concern for all countries as it undermines the public finance and can lead to a very inefficient resource allocation. Therefore it is of great interest to develop models that can help targeting the units (firms or individuals) at risk of noncompliance. Our focus is on a particular facet of tax evasion: undeclared work. We use an ad hoc data set by linking information on the inspections carried out in Italy in 2005 by the National Institute of Social Security on building and construction companies and a vast set of firms characteristics. In the analysis there are two methodological issues: the non-randomness of the inspections and the very low fraction of inspected firms in the population (around 5%). We propose a method that combine the Heckman-type model with sample selection and the response-based sampling. We find that the likelihood function is a weighted version of the Heckman model, where the weights take into account the sampling scheme.